tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  */
  16
  17 #include "util/record.h"
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "util/rlimit.h"
  23 #include "builtin.h"
  24 #include "util/cgroup.h"
  25 #include "util/color.h"
  26 #include "util/config.h"
  27 #include "util/debug.h"
  28 #include "util/dso.h"
  29 #include "util/env.h"
  30 #include "util/event.h"
  31 #include "util/evsel.h"
  32 #include "util/evsel_fprintf.h"
  33 #include "util/synthetic-events.h"
  34 #include "util/evlist.h"
  35 #include "util/evswitch.h"
  36 #include "util/mmap.h"
  37 #include <subcmd/pager.h>
  38 #include <subcmd/exec-cmd.h>
  39 #include "util/machine.h"
  40 #include "util/map.h"
  41 #include "util/symbol.h"
  42 #include "util/path.h"
  43 #include "util/session.h"
  44 #include "util/thread.h"
  45 #include <subcmd/parse-options.h>
  46 #include "util/strlist.h"
  47 #include "util/intlist.h"
  48 #include "util/thread_map.h"
  49 #include "util/stat.h"
  50 #include "util/tool.h"
  51 #include "util/util.h"
  52 #include "trace/beauty/beauty.h"
  53 #include "trace-event.h"
  54 #include "util/parse-events.h"
  55 #include "util/bpf-loader.h"
  56 #include "callchain.h"
  57 #include "print_binary.h"
  58 #include "string2.h"
  59 #include "syscalltbl.h"
  60 #include "rb_resort.h"
  61 #include "../perf.h"
  62
  63 #include <errno.h>
  64 #include <inttypes.h>
  65 #include <poll.h>
  66 #include <signal.h>
  67 #include <stdlib.h>
  68 #include <string.h>
  69 #include <linux/err.h>
  70 #include <linux/filter.h>
  71 #include <linux/kernel.h>
  72 #include <linux/random.h>
  73 #include <linux/stringify.h>
  74 #include <linux/time64.h>
  75 #include <linux/zalloc.h>
  76 #include <fcntl.h>
  77 #include <sys/sysmacros.h>
  78
  79 #include <linux/ctype.h>
  80 #include <perf/mmap.h>
  81
  82 #ifndef O_CLOEXEC
  83 # define O_CLOEXEC              02000000
  84 #endif
  85
  86 #ifndef F_LINUX_SPECIFIC_BASE
  87 # define F_LINUX_SPECIFIC_BASE  1024
  88 #endif
  89
  90 /*
  91  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
  92  */
  93 struct syscall_arg_fmt {
  94         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
  95         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
  96         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
  97         void       *parm;
  98         const char *name;
  99         u16        nr_entries; // for arrays
 100         bool       show_zero;
 101 };
 102
 103 struct syscall_fmt {
 104         const char *name;
 105         const char *alias;
 106         struct {
 107                 const char *sys_enter,
 108                            *sys_exit;
 109         }          bpf_prog_name;
 110         struct syscall_arg_fmt arg[6];
 111         u8         nr_args;
 112         bool       errpid;
 113         bool       timeout;
 114         bool       hexret;
 115 };
 116
 117 struct trace {
 118         struct perf_tool        tool;
 119         struct syscalltbl       *sctbl;
 120         struct {
 121                 struct syscall  *table;
 122                 struct bpf_map  *map;
 123                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 124                         struct bpf_map  *sys_enter,
 125                                         *sys_exit;
 126                 }               prog_array;
 127                 struct {
 128                         struct evsel *sys_enter,
 129                                           *sys_exit,
 130                                           *augmented;
 131                 }               events;
 132                 struct bpf_program *unaugmented_prog;
 133         } syscalls;
 134         struct {
 135                 struct bpf_map *map;
 136         } dump;
 137         struct record_opts      opts;
 138         struct evlist   *evlist;
 139         struct machine          *host;
 140         struct thread           *current;
 141         struct bpf_object       *bpf_obj;
 142         struct cgroup           *cgroup;
 143         u64                     base_time;
 144         FILE                    *output;
 145         unsigned long           nr_events;
 146         unsigned long           nr_events_printed;
 147         unsigned long           max_events;
 148         struct evswitch         evswitch;
 149         struct strlist          *ev_qualifier;
 150         struct {
 151                 size_t          nr;
 152                 int             *entries;
 153         }                       ev_qualifier_ids;
 154         struct {
 155                 size_t          nr;
 156                 pid_t           *entries;
 157                 struct bpf_map  *map;
 158         }                       filter_pids;
 159         double                  duration_filter;
 160         double                  runtime_ms;
 161         struct {
 162                 u64             vfs_getname,
 163                                 proc_getname;
 164         } stats;
 165         unsigned int            max_stack;
 166         unsigned int            min_stack;
 167         int                     raw_augmented_syscalls_args_size;
 168         bool                    raw_augmented_syscalls;
 169         bool                    fd_path_disabled;
 170         bool                    sort_events;
 171         bool                    not_ev_qualifier;
 172         bool                    live;
 173         bool                    full_time;
 174         bool                    sched;
 175         bool                    multiple_threads;
 176         bool                    summary;
 177         bool                    summary_only;
 178         bool                    errno_summary;
 179         bool                    failure_only;
 180         bool                    show_comm;
 181         bool                    print_sample;
 182         bool                    show_tool_stats;
 183         bool                    trace_syscalls;
 184         bool                    libtraceevent_print;
 185         bool                    kernel_syscallchains;
 186         s16                     args_alignment;
 187         bool                    show_tstamp;
 188         bool                    show_duration;
 189         bool                    show_zeros;
 190         bool                    show_arg_names;
 191         bool                    show_string_prefix;
 192         bool                    force;
 193         bool                    vfs_getname;
 194         int                     trace_pgfaults;
 195         char                    *perfconfig_events;
 196         struct {
 197                 struct ordered_events   data;
 198                 u64                     last;
 199         } oe;
 200 };
 201
 202 struct tp_field {
 203         int offset;
 204         union {
 205                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 206                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 207         };
 208 };
 209
 210 #define TP_UINT_FIELD(bits) \
 211 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 212 { \
 213         u##bits value; \
 214         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 215         return value;  \
 216 }
 217
 218 TP_UINT_FIELD(8);
 219 TP_UINT_FIELD(16);
 220 TP_UINT_FIELD(32);
 221 TP_UINT_FIELD(64);
 222
 223 #define TP_UINT_FIELD__SWAPPED(bits) \
 224 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 225 { \
 226         u##bits value; \
 227         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 228         return bswap_##bits(value);\
 229 }
 230
 231 TP_UINT_FIELD__SWAPPED(16);
 232 TP_UINT_FIELD__SWAPPED(32);
 233 TP_UINT_FIELD__SWAPPED(64);
 234
 235 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 236 {
 237         field->offset = offset;
 238
 239         switch (size) {
 240         case 1:
 241                 field->integer = tp_field__u8;
 242                 break;
 243         case 2:
 244                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 245                 break;
 246         case 4:
 247                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 248                 break;
 249         case 8:
 250                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 251                 break;
 252         default:
 253                 return -1;
 254         }
 255
 256         return 0;
 257 }
 258
 259 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 260 {
 261         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 262 }
 263
 264 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 265 {
 266         return sample->raw_data + field->offset;
 267 }
 268
 269 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 270 {
 271         field->offset = offset;
 272         field->pointer = tp_field__ptr;
 273         return 0;
 274 }
 275
 276 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 277 {
 278         return __tp_field__init_ptr(field, format_field->offset);
 279 }
 280
 281 struct syscall_tp {
 282         struct tp_field id;
 283         union {
 284                 struct tp_field args, ret;
 285         };
 286 };
 287
 288 /*
 289  * The evsel->priv as used by 'perf trace'
 290  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 291  * fmt: for all the other tracepoints
 292  */
 293 struct evsel_trace {
 294         struct syscall_tp       sc;
 295         struct syscall_arg_fmt  *fmt;
 296 };
 297
 298 static struct evsel_trace *evsel_trace__new(void)
 299 {
 300         return zalloc(sizeof(struct evsel_trace));
 301 }
 302
 303 static void evsel_trace__delete(struct evsel_trace *et)
 304 {
 305         if (et == NULL)
 306                 return;
 307
 308         zfree(&et->fmt);
 309         free(et);
 310 }
 311
 312 /*
 313  * Used with raw_syscalls:sys_{enter,exit} and with the
 314  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 315  */
 316 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 317 {
 318         struct evsel_trace *et = evsel->priv;
 319
 320         return &et->sc;
 321 }
 322
 323 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 324 {
 325         if (evsel->priv == NULL) {
 326                 evsel->priv = evsel_trace__new();
 327                 if (evsel->priv == NULL)
 328                         return NULL;
 329         }
 330
 331         return __evsel__syscall_tp(evsel);
 332 }
 333
 334 /*
 335  * Used with all the other tracepoints.
 336  */
 337 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 338 {
 339         struct evsel_trace *et = evsel->priv;
 340
 341         return et->fmt;
 342 }
 343
 344 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 345 {
 346         struct evsel_trace *et = evsel->priv;
 347
 348         if (evsel->priv == NULL) {
 349                 et = evsel->priv = evsel_trace__new();
 350
 351                 if (et == NULL)
 352                         return NULL;
 353         }
 354
 355         if (et->fmt == NULL) {
 356                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 357                 if (et->fmt == NULL)
 358                         goto out_delete;
 359         }
 360
 361         return __evsel__syscall_arg_fmt(evsel);
 362
 363 out_delete:
 364         evsel_trace__delete(evsel->priv);
 365         evsel->priv = NULL;
 366         return NULL;
 367 }
 368
 369 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 370 {
 371         struct tep_format_field *format_field = evsel__field(evsel, name);
 372
 373         if (format_field == NULL)
 374                 return -1;
 375
 376         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 377 }
 378
 379 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 380         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 381            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 382
 383 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 384 {
 385         struct tep_format_field *format_field = evsel__field(evsel, name);
 386
 387         if (format_field == NULL)
 388                 return -1;
 389
 390         return tp_field__init_ptr(field, format_field);
 391 }
 392
 393 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 394         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 395            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 396
 397 static void evsel__delete_priv(struct evsel *evsel)
 398 {
 399         zfree(&evsel->priv);
 400         evsel__delete(evsel);
 401 }
 402
 403 static int evsel__init_syscall_tp(struct evsel *evsel)
 404 {
 405         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 406
 407         if (sc != NULL) {
 408                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 409                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 410                         return -ENOENT;
 411                 return 0;
 412         }
 413
 414         return -ENOMEM;
 415 }
 416
 417 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 418 {
 419         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 420
 421         if (sc != NULL) {
 422                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 423                 if (syscall_id == NULL)
 424                         syscall_id = evsel__field(tp, "__syscall_nr");
 425                 if (syscall_id == NULL ||
 426                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 427                         return -EINVAL;
 428
 429                 return 0;
 430         }
 431
 432         return -ENOMEM;
 433 }
 434
 435 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 436 {
 437         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 438
 439         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 440 }
 441
 442 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 443 {
 444         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 445
 446         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 447 }
 448
 449 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 450 {
 451         if (evsel__syscall_tp(evsel) != NULL) {
 452                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 453                         return -ENOENT;
 454
 455                 evsel->handler = handler;
 456                 return 0;
 457         }
 458
 459         return -ENOMEM;
 460 }
 461
 462 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 463 {
 464         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 465
 466         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 467         if (IS_ERR(evsel))
 468                 evsel = evsel__newtp("syscalls", direction);
 469
 470         if (IS_ERR(evsel))
 471                 return NULL;
 472
 473         if (evsel__init_raw_syscall_tp(evsel, handler))
 474                 goto out_delete;
 475
 476         return evsel;
 477
 478 out_delete:
 479         evsel__delete_priv(evsel);
 480         return NULL;
 481 }
 482
 483 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 484         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 485            fields->name.integer(&fields->name, sample); })
 486
 487 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 488         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 489            fields->name.pointer(&fields->name, sample); })
 490
 491 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 492 {
 493         int idx = val - sa->offset;
 494
 495         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 496                 size_t printed = scnprintf(bf, size, intfmt, val);
 497                 if (show_suffix)
 498                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 499                 return printed;
 500         }
 501
 502         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 503 }
 504
 505 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 506 {
 507         int idx = val - sa->offset;
 508
 509         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 510                 size_t printed = scnprintf(bf, size, intfmt, val);
 511                 if (show_prefix)
 512                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 513                 return printed;
 514         }
 515
 516         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 517 }
 518
 519 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 520                                                 const char *intfmt,
 521                                                 struct syscall_arg *arg)
 522 {
 523         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 524 }
 525
 526 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 527                                               struct syscall_arg *arg)
 528 {
 529         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 530 }
 531
 532 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 533
 534 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 535 {
 536         return strarray__strtoul(arg->parm, bf, size, ret);
 537 }
 538
 539 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 540 {
 541         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 542 }
 543
 544 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 545 {
 546         return strarrays__strtoul(arg->parm, bf, size, ret);
 547 }
 548
 549 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 550 {
 551         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 552 }
 553
 554 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 555 {
 556         size_t printed;
 557         int i;
 558
 559         for (i = 0; i < sas->nr_entries; ++i) {
 560                 struct strarray *sa = sas->entries[i];
 561                 int idx = val - sa->offset;
 562
 563                 if (idx >= 0 && idx < sa->nr_entries) {
 564                         if (sa->entries[idx] == NULL)
 565                                 break;
 566                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 567                 }
 568         }
 569
 570         printed = scnprintf(bf, size, intfmt, val);
 571         if (show_prefix)
 572                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 573         return printed;
 574 }
 575
 576 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 577 {
 578         int i;
 579
 580         for (i = 0; i < sa->nr_entries; ++i) {
 581                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 582                         *ret = sa->offset + i;
 583                         return true;
 584                 }
 585         }
 586
 587         return false;
 588 }
 589
 590 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 591 {
 592         u64 val = 0;
 593         char *tok = bf, *sep, *end;
 594
 595         *ret = 0;
 596
 597         while (size != 0) {
 598                 int toklen = size;
 599
 600                 sep = memchr(tok, '|', size);
 601                 if (sep != NULL) {
 602                         size -= sep - tok + 1;
 603
 604                         end = sep - 1;
 605                         while (end > tok && isspace(*end))
 606                                 --end;
 607
 608                         toklen = end - tok + 1;
 609                 }
 610
 611                 while (isspace(*tok))
 612                         ++tok;
 613
 614                 if (isalpha(*tok) || *tok == '_') {
 615                         if (!strarray__strtoul(sa, tok, toklen, &val))
 616                                 return false;
 617                 } else {
 618                         bool is_hexa = tok[0] == 0 && (tok[1] = 'x' || tok[1] == 'X');
 619
 620                         val = strtoul(tok, NULL, is_hexa ? 16 : 0);
 621                 }
 622
 623                 *ret |= (1 << (val - 1));
 624
 625                 if (sep == NULL)
 626                         break;
 627                 tok = sep + 1;
 628         }
 629
 630         return true;
 631 }
 632
 633 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 634 {
 635         int i;
 636
 637         for (i = 0; i < sas->nr_entries; ++i) {
 638                 struct strarray *sa = sas->entries[i];
 639
 640                 if (strarray__strtoul(sa, bf, size, ret))
 641                         return true;
 642         }
 643
 644         return false;
 645 }
 646
 647 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 648                                         struct syscall_arg *arg)
 649 {
 650         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 651 }
 652
 653 #ifndef AT_FDCWD
 654 #define AT_FDCWD        -100
 655 #endif
 656
 657 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 658                                            struct syscall_arg *arg)
 659 {
 660         int fd = arg->val;
 661         const char *prefix = "AT_FD";
 662
 663         if (fd == AT_FDCWD)
 664                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 665
 666         return syscall_arg__scnprintf_fd(bf, size, arg);
 667 }
 668
 669 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 670
 671 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 672                                               struct syscall_arg *arg);
 673
 674 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 675
 676 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 677 {
 678         return scnprintf(bf, size, "%#lx", arg->val);
 679 }
 680
 681 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 682 {
 683         if (arg->val == 0)
 684                 return scnprintf(bf, size, "NULL");
 685         return syscall_arg__scnprintf_hex(bf, size, arg);
 686 }
 687
 688 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 689 {
 690         return scnprintf(bf, size, "%d", arg->val);
 691 }
 692
 693 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 694 {
 695         return scnprintf(bf, size, "%ld", arg->val);
 696 }
 697
 698 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 699 {
 700         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 701         //     fill missing comms using thread__set_comm()...
 702         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 703         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 704 }
 705
 706 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 707
 708 static const char *bpf_cmd[] = {
 709         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 710         "MAP_GET_NEXT_KEY", "PROG_LOAD", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 711         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 712         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 713         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 714         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 715         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 716         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 717         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 718         "LINK_DETACH", "PROG_BIND_MAP",
 719 };
 720 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 721
 722 static const char *fsmount_flags[] = {
 723         [1] = "CLOEXEC",
 724 };
 725 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 726
 727 #include "trace/beauty/generated/fsconfig_arrays.c"
 728
 729 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 730
 731 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 732 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 733
 734 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 735 static DEFINE_STRARRAY(itimers, "ITIMER_");
 736
 737 static const char *keyctl_options[] = {
 738         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 739         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 740         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 741         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 742         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 743 };
 744 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 745
 746 static const char *whences[] = { "SET", "CUR", "END",
 747 #ifdef SEEK_DATA
 748 "DATA",
 749 #endif
 750 #ifdef SEEK_HOLE
 751 "HOLE",
 752 #endif
 753 };
 754 static DEFINE_STRARRAY(whences, "SEEK_");
 755
 756 static const char *fcntl_cmds[] = {
 757         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 758         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 759         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 760         "GETOWNER_UIDS",
 761 };
 762 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 763
 764 static const char *fcntl_linux_specific_cmds[] = {
 765         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 766         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 767         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 768 };
 769
 770 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 771
 772 static struct strarray *fcntl_cmds_arrays[] = {
 773         &strarray__fcntl_cmds,
 774         &strarray__fcntl_linux_specific_cmds,
 775 };
 776
 777 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 778
 779 static const char *rlimit_resources[] = {
 780         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 781         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 782         "RTTIME",
 783 };
 784 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 785
 786 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 787 static DEFINE_STRARRAY(sighow, "SIG_");
 788
 789 static const char *clockid[] = {
 790         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 791         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 792         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 793 };
 794 static DEFINE_STRARRAY(clockid, "CLOCK_");
 795
 796 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 797                                                  struct syscall_arg *arg)
 798 {
 799         bool show_prefix = arg->show_string_prefix;
 800         const char *suffix = "_OK";
 801         size_t printed = 0;
 802         int mode = arg->val;
 803
 804         if (mode == F_OK) /* 0 */
 805                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 806 #define P_MODE(n) \
 807         if (mode & n##_OK) { \
 808                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 809                 mode &= ~n##_OK; \
 810         }
 811
 812         P_MODE(R);
 813         P_MODE(W);
 814         P_MODE(X);
 815 #undef P_MODE
 816
 817         if (mode)
 818                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 819
 820         return printed;
 821 }
 822
 823 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 824
 825 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 826                                               struct syscall_arg *arg);
 827
 828 #define SCA_FILENAME syscall_arg__scnprintf_filename
 829
 830 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 831                                                 struct syscall_arg *arg)
 832 {
 833         bool show_prefix = arg->show_string_prefix;
 834         const char *prefix = "O_";
 835         int printed = 0, flags = arg->val;
 836
 837 #define P_FLAG(n) \
 838         if (flags & O_##n) { \
 839                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 840                 flags &= ~O_##n; \
 841         }
 842
 843         P_FLAG(CLOEXEC);
 844         P_FLAG(NONBLOCK);
 845 #undef P_FLAG
 846
 847         if (flags)
 848                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 849
 850         return printed;
 851 }
 852
 853 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 854
 855 #ifndef GRND_NONBLOCK
 856 #define GRND_NONBLOCK   0x0001
 857 #endif
 858 #ifndef GRND_RANDOM
 859 #define GRND_RANDOM     0x0002
 860 #endif
 861
 862 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 863                                                    struct syscall_arg *arg)
 864 {
 865         bool show_prefix = arg->show_string_prefix;
 866         const char *prefix = "GRND_";
 867         int printed = 0, flags = arg->val;
 868
 869 #define P_FLAG(n) \
 870         if (flags & GRND_##n) { \
 871                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 872                 flags &= ~GRND_##n; \
 873         }
 874
 875         P_FLAG(RANDOM);
 876         P_FLAG(NONBLOCK);
 877 #undef P_FLAG
 878
 879         if (flags)
 880                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 881
 882         return printed;
 883 }
 884
 885 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 886
 887 #define STRARRAY(name, array) \
 888           { .scnprintf  = SCA_STRARRAY, \
 889             .strtoul    = STUL_STRARRAY, \
 890             .parm       = &strarray__##array, }
 891
 892 #define STRARRAY_FLAGS(name, array) \
 893           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 894             .strtoul    = STUL_STRARRAY_FLAGS, \
 895             .parm       = &strarray__##array, }
 896
 897 #include "trace/beauty/arch_errno_names.c"
 898 #include "trace/beauty/eventfd.c"
 899 #include "trace/beauty/futex_op.c"
 900 #include "trace/beauty/futex_val3.c"
 901 #include "trace/beauty/mmap.c"
 902 #include "trace/beauty/mode_t.c"
 903 #include "trace/beauty/msg_flags.c"
 904 #include "trace/beauty/open_flags.c"
 905 #include "trace/beauty/perf_event_open.c"
 906 #include "trace/beauty/pid.c"
 907 #include "trace/beauty/sched_policy.c"
 908 #include "trace/beauty/seccomp.c"
 909 #include "trace/beauty/signum.c"
 910 #include "trace/beauty/socket_type.c"
 911 #include "trace/beauty/waitid_options.c"
 912
 913 static struct syscall_fmt syscall_fmts[] = {
 914         { .name     = "access",
 915           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 916         { .name     = "arch_prctl",
 917           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 918                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 919         { .name     = "bind",
 920           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 921                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 922                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 923         { .name     = "bpf",
 924           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 925         { .name     = "brk",        .hexret = true,
 926           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 927         { .name     = "clock_gettime",
 928           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 929         { .name     = "clone",      .errpid = true, .nr_args = 5,
 930           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 931                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 932                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 933                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 934                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 935         { .name     = "close",
 936           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 937         { .name     = "connect",
 938           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 939                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 940                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 941         { .name     = "epoll_ctl",
 942           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 943         { .name     = "eventfd2",
 944           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 945         { .name     = "fchmodat",
 946           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 947         { .name     = "fchownat",
 948           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 949         { .name     = "fcntl",
 950           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 951                            .strtoul   = STUL_STRARRAYS,
 952                            .parm      = &strarrays__fcntl_cmds_arrays,
 953                            .show_zero = true, },
 954                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 955         { .name     = "flock",
 956           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 957         { .name     = "fsconfig",
 958           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 959         { .name     = "fsmount",
 960           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 961                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 962         { .name     = "fspick",
 963           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 964                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 965                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 966         { .name     = "fstat", .alias = "newfstat", },
 967         { .name     = "fstatat", .alias = "newfstatat", },
 968         { .name     = "futex",
 969           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 970                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 971         { .name     = "futimesat",
 972           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 973         { .name     = "getitimer",
 974           .arg = { [0] = STRARRAY(which, itimers), }, },
 975         { .name     = "getpid",     .errpid = true, },
 976         { .name     = "getpgid",    .errpid = true, },
 977         { .name     = "getppid",    .errpid = true, },
 978         { .name     = "getrandom",
 979           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 980         { .name     = "getrlimit",
 981           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 982         { .name     = "getsockopt",
 983           .arg = { [1] = STRARRAY(level, socket_level), }, },
 984         { .name     = "gettid",     .errpid = true, },
 985         { .name     = "ioctl",
 986           .arg = {
 987 #if defined(__i386__) || defined(__x86_64__)
 988 /*
 989  * FIXME: Make this available to all arches.
 990  */
 991                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 992                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 993 #else
 994                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 995 #endif
 996         { .name     = "kcmp",       .nr_args = 5,
 997           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 998                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 999                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
1000                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
1001                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1002         { .name     = "keyctl",
1003           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1004         { .name     = "kill",
1005           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1006         { .name     = "linkat",
1007           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1008         { .name     = "lseek",
1009           .arg = { [2] = STRARRAY(whence, whences), }, },
1010         { .name     = "lstat", .alias = "newlstat", },
1011         { .name     = "madvise",
1012           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1013                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1014         { .name     = "mkdirat",
1015           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1016         { .name     = "mknodat",
1017           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1018         { .name     = "mmap",       .hexret = true,
1019 /* The standard mmap maps to old_mmap on s390x */
1020 #if defined(__s390x__)
1021         .alias = "old_mmap",
1022 #endif
1023           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1024                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1025                            .strtoul   = STUL_STRARRAY_FLAGS,
1026                            .parm      = &strarray__mmap_flags, },
1027                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1028         { .name     = "mount",
1029           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1030                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1031                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1032         { .name     = "move_mount",
1033           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1034                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1035                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1036                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1037                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1038         { .name     = "mprotect",
1039           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1040                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1041         { .name     = "mq_unlink",
1042           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1043         { .name     = "mremap",     .hexret = true,
1044           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1045         { .name     = "name_to_handle_at",
1046           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1047         { .name     = "newfstatat",
1048           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1049         { .name     = "open",
1050           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1051         { .name     = "open_by_handle_at",
1052           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1053                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1054         { .name     = "openat",
1055           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1056                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1057         { .name     = "perf_event_open",
1058           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1059                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1060                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1061         { .name     = "pipe2",
1062           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1063         { .name     = "pkey_alloc",
1064           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1065         { .name     = "pkey_free",
1066           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1067         { .name     = "pkey_mprotect",
1068           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1069                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1070                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1071         { .name     = "poll", .timeout = true, },
1072         { .name     = "ppoll", .timeout = true, },
1073         { .name     = "prctl",
1074           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1075                            .strtoul   = STUL_STRARRAY,
1076                            .parm      = &strarray__prctl_options, },
1077                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1078                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1079         { .name     = "pread", .alias = "pread64", },
1080         { .name     = "preadv", .alias = "pread", },
1081         { .name     = "prlimit64",
1082           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1083         { .name     = "pwrite", .alias = "pwrite64", },
1084         { .name     = "readlinkat",
1085           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1086         { .name     = "recvfrom",
1087           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1088         { .name     = "recvmmsg",
1089           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1090         { .name     = "recvmsg",
1091           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1092         { .name     = "renameat",
1093           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1094                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1095         { .name     = "renameat2",
1096           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1097                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1098                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1099         { .name     = "rt_sigaction",
1100           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1101         { .name     = "rt_sigprocmask",
1102           .arg = { [0] = STRARRAY(how, sighow), }, },
1103         { .name     = "rt_sigqueueinfo",
1104           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1105         { .name     = "rt_tgsigqueueinfo",
1106           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1107         { .name     = "sched_setscheduler",
1108           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1109         { .name     = "seccomp",
1110           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1111                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1112         { .name     = "select", .timeout = true, },
1113         { .name     = "sendfile", .alias = "sendfile64", },
1114         { .name     = "sendmmsg",
1115           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1116         { .name     = "sendmsg",
1117           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1118         { .name     = "sendto",
1119           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1120                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1121         { .name     = "set_tid_address", .errpid = true, },
1122         { .name     = "setitimer",
1123           .arg = { [0] = STRARRAY(which, itimers), }, },
1124         { .name     = "setrlimit",
1125           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1126         { .name     = "setsockopt",
1127           .arg = { [1] = STRARRAY(level, socket_level), }, },
1128         { .name     = "socket",
1129           .arg = { [0] = STRARRAY(family, socket_families),
1130                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1131                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1132         { .name     = "socketpair",
1133           .arg = { [0] = STRARRAY(family, socket_families),
1134                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1135                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1136         { .name     = "stat", .alias = "newstat", },
1137         { .name     = "statx",
1138           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1139                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1140                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1141         { .name     = "swapoff",
1142           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1143         { .name     = "swapon",
1144           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1145         { .name     = "symlinkat",
1146           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1147         { .name     = "sync_file_range",
1148           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1149         { .name     = "tgkill",
1150           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1151         { .name     = "tkill",
1152           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1153         { .name     = "umount2", .alias = "umount",
1154           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1155         { .name     = "uname", .alias = "newuname", },
1156         { .name     = "unlinkat",
1157           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1158         { .name     = "utimensat",
1159           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1160         { .name     = "wait4",      .errpid = true,
1161           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1162         { .name     = "waitid",     .errpid = true,
1163           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1164 };
1165
1166 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1167 {
1168         const struct syscall_fmt *fmt = fmtp;
1169         return strcmp(name, fmt->name);
1170 }
1171
1172 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1173 {
1174         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1175 }
1176
1177 static struct syscall_fmt *syscall_fmt__find(const char *name)
1178 {
1179         const int nmemb = ARRAY_SIZE(syscall_fmts);
1180         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1181 }
1182
1183 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1184 {
1185         int i;
1186
1187         for (i = 0; i < nmemb; ++i) {
1188                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1189                         return &fmts[i];
1190         }
1191
1192         return NULL;
1193 }
1194
1195 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1196 {
1197         const int nmemb = ARRAY_SIZE(syscall_fmts);
1198         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1199 }
1200
1201 /*
1202  * is_exit: is this "exit" or "exit_group"?
1203  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1204  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1205  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1206  */
1207 struct syscall {
1208         struct tep_event    *tp_format;
1209         int                 nr_args;
1210         int                 args_size;
1211         struct {
1212                 struct bpf_program *sys_enter,
1213                                    *sys_exit;
1214         }                   bpf_prog;
1215         bool                is_exit;
1216         bool                is_open;
1217         bool                nonexistent;
1218         struct tep_format_field *args;
1219         const char          *name;
1220         struct syscall_fmt  *fmt;
1221         struct syscall_arg_fmt *arg_fmt;
1222 };
1223
1224 /*
1225  * Must match what is in the BPF program:
1226  *
1227  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1228  */
1229 struct bpf_map_syscall_entry {
1230         bool    enabled;
1231         u16     string_args_len[6];
1232 };
1233
1234 /*
1235  * We need to have this 'calculated' boolean because in some cases we really
1236  * don't know what is the duration of a syscall, for instance, when we start
1237  * a session and some threads are waiting for a syscall to finish, say 'poll',
1238  * in which case all we can do is to print "( ? ) for duration and for the
1239  * start timestamp.
1240  */
1241 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1242 {
1243         double duration = (double)t / NSEC_PER_MSEC;
1244         size_t printed = fprintf(fp, "(");
1245
1246         if (!calculated)
1247                 printed += fprintf(fp, "         ");
1248         else if (duration >= 1.0)
1249                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1250         else if (duration >= 0.01)
1251                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1252         else
1253                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1254         return printed + fprintf(fp, "): ");
1255 }
1256
1257 /**
1258  * filename.ptr: The filename char pointer that will be vfs_getname'd
1259  * filename.entry_str_pos: Where to insert the string translated from
1260  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1261  * ret_scnprintf: syscall args may set this to a different syscall return
1262  *                formatter, for instance, fcntl may return fds, file flags, etc.
1263  */
1264 struct thread_trace {
1265         u64               entry_time;
1266         bool              entry_pending;
1267         unsigned long     nr_events;
1268         unsigned long     pfmaj, pfmin;
1269         char              *entry_str;
1270         double            runtime_ms;
1271         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1272         struct {
1273                 unsigned long ptr;
1274                 short int     entry_str_pos;
1275                 bool          pending_open;
1276                 unsigned int  namelen;
1277                 char          *name;
1278         } filename;
1279         struct {
1280                 int           max;
1281                 struct file   *table;
1282         } files;
1283
1284         struct intlist *syscall_stats;
1285 };
1286
1287 static struct thread_trace *thread_trace__new(void)
1288 {
1289         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1290
1291         if (ttrace) {
1292                 ttrace->files.max = -1;
1293                 ttrace->syscall_stats = intlist__new(NULL);
1294         }
1295
1296         return ttrace;
1297 }
1298
1299 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1300 {
1301         struct thread_trace *ttrace;
1302
1303         if (thread == NULL)
1304                 goto fail;
1305
1306         if (thread__priv(thread) == NULL)
1307                 thread__set_priv(thread, thread_trace__new());
1308
1309         if (thread__priv(thread) == NULL)
1310                 goto fail;
1311
1312         ttrace = thread__priv(thread);
1313         ++ttrace->nr_events;
1314
1315         return ttrace;
1316 fail:
1317         color_fprintf(fp, PERF_COLOR_RED,
1318                       "WARNING: not enough memory, dropping samples!\n");
1319         return NULL;
1320 }
1321
1322
1323 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1324                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1325 {
1326         struct thread_trace *ttrace = thread__priv(arg->thread);
1327
1328         ttrace->ret_scnprintf = ret_scnprintf;
1329 }
1330
1331 #define TRACE_PFMAJ             (1 << 0)
1332 #define TRACE_PFMIN             (1 << 1)
1333
1334 static const size_t trace__entry_str_size = 2048;
1335
1336 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1337 {
1338         if (fd < 0)
1339                 return NULL;
1340
1341         if (fd > ttrace->files.max) {
1342                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1343
1344                 if (nfiles == NULL)
1345                         return NULL;
1346
1347                 if (ttrace->files.max != -1) {
1348                         memset(nfiles + ttrace->files.max + 1, 0,
1349                                (fd - ttrace->files.max) * sizeof(struct file));
1350                 } else {
1351                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1352                 }
1353
1354                 ttrace->files.table = nfiles;
1355                 ttrace->files.max   = fd;
1356         }
1357
1358         return ttrace->files.table + fd;
1359 }
1360
1361 struct file *thread__files_entry(struct thread *thread, int fd)
1362 {
1363         return thread_trace__files_entry(thread__priv(thread), fd);
1364 }
1365
1366 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1367 {
1368         struct thread_trace *ttrace = thread__priv(thread);
1369         struct file *file = thread_trace__files_entry(ttrace, fd);
1370
1371         if (file != NULL) {
1372                 struct stat st;
1373                 if (stat(pathname, &st) == 0)
1374                         file->dev_maj = major(st.st_rdev);
1375                 file->pathname = strdup(pathname);
1376                 if (file->pathname)
1377                         return 0;
1378         }
1379
1380         return -1;
1381 }
1382
1383 static int thread__read_fd_path(struct thread *thread, int fd)
1384 {
1385         char linkname[PATH_MAX], pathname[PATH_MAX];
1386         struct stat st;
1387         int ret;
1388
1389         if (thread->pid_ == thread->tid) {
1390                 scnprintf(linkname, sizeof(linkname),
1391                           "/proc/%d/fd/%d", thread->pid_, fd);
1392         } else {
1393                 scnprintf(linkname, sizeof(linkname),
1394                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1395         }
1396
1397         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1398                 return -1;
1399
1400         ret = readlink(linkname, pathname, sizeof(pathname));
1401
1402         if (ret < 0 || ret > st.st_size)
1403                 return -1;
1404
1405         pathname[ret] = '\0';
1406         return trace__set_fd_pathname(thread, fd, pathname);
1407 }
1408
1409 static const char *thread__fd_path(struct thread *thread, int fd,
1410                                    struct trace *trace)
1411 {
1412         struct thread_trace *ttrace = thread__priv(thread);
1413
1414         if (ttrace == NULL || trace->fd_path_disabled)
1415                 return NULL;
1416
1417         if (fd < 0)
1418                 return NULL;
1419
1420         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1421                 if (!trace->live)
1422                         return NULL;
1423                 ++trace->stats.proc_getname;
1424                 if (thread__read_fd_path(thread, fd))
1425                         return NULL;
1426         }
1427
1428         return ttrace->files.table[fd].pathname;
1429 }
1430
1431 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1432 {
1433         int fd = arg->val;
1434         size_t printed = scnprintf(bf, size, "%d", fd);
1435         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1436
1437         if (path)
1438                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1439
1440         return printed;
1441 }
1442
1443 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1444 {
1445         size_t printed = scnprintf(bf, size, "%d", fd);
1446         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1447
1448         if (thread) {
1449                 const char *path = thread__fd_path(thread, fd, trace);
1450
1451                 if (path)
1452                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1453
1454                 thread__put(thread);
1455         }
1456
1457         return printed;
1458 }
1459
1460 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1461                                               struct syscall_arg *arg)
1462 {
1463         int fd = arg->val;
1464         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1465         struct thread_trace *ttrace = thread__priv(arg->thread);
1466
1467         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1468                 zfree(&ttrace->files.table[fd].pathname);
1469
1470         return printed;
1471 }
1472
1473 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1474                                      unsigned long ptr)
1475 {
1476         struct thread_trace *ttrace = thread__priv(thread);
1477
1478         ttrace->filename.ptr = ptr;
1479         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1480 }
1481
1482 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1483 {
1484         struct augmented_arg *augmented_arg = arg->augmented.args;
1485         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1486         /*
1487          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1488          * we would have two strings, each prefixed by its size.
1489          */
1490         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1491
1492         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1493         arg->augmented.size -= consumed;
1494
1495         return printed;
1496 }
1497
1498 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1499                                               struct syscall_arg *arg)
1500 {
1501         unsigned long ptr = arg->val;
1502
1503         if (arg->augmented.args)
1504                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1505
1506         if (!arg->trace->vfs_getname)
1507                 return scnprintf(bf, size, "%#x", ptr);
1508
1509         thread__set_filename_pos(arg->thread, bf, ptr);
1510         return 0;
1511 }
1512
1513 static bool trace__filter_duration(struct trace *trace, double t)
1514 {
1515         return t < (trace->duration_filter * NSEC_PER_MSEC);
1516 }
1517
1518 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1519 {
1520         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1521
1522         return fprintf(fp, "%10.3f ", ts);
1523 }
1524
1525 /*
1526  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1527  * using ttrace->entry_time for a thread that receives a sys_exit without
1528  * first having received a sys_enter ("poll" issued before tracing session
1529  * starts, lost sys_enter exit due to ring buffer overflow).
1530  */
1531 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1532 {
1533         if (tstamp > 0)
1534                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1535
1536         return fprintf(fp, "         ? ");
1537 }
1538
1539 static bool done = false;
1540 static bool interrupted = false;
1541
1542 static void sig_handler(int sig)
1543 {
1544         done = true;
1545         interrupted = sig == SIGINT;
1546 }
1547
1548 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1549 {
1550         size_t printed = 0;
1551
1552         if (trace->multiple_threads) {
1553                 if (trace->show_comm)
1554                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1555                 printed += fprintf(fp, "%d ", thread->tid);
1556         }
1557
1558         return printed;
1559 }
1560
1561 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1562                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1563 {
1564         size_t printed = 0;
1565
1566         if (trace->show_tstamp)
1567                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1568         if (trace->show_duration)
1569                 printed += fprintf_duration(duration, duration_calculated, fp);
1570         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1571 }
1572
1573 static int trace__process_event(struct trace *trace, struct machine *machine,
1574                                 union perf_event *event, struct perf_sample *sample)
1575 {
1576         int ret = 0;
1577
1578         switch (event->header.type) {
1579         case PERF_RECORD_LOST:
1580                 color_fprintf(trace->output, PERF_COLOR_RED,
1581                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1582                 ret = machine__process_lost_event(machine, event, sample);
1583                 break;
1584         default:
1585                 ret = machine__process_event(machine, event, sample);
1586                 break;
1587         }
1588
1589         return ret;
1590 }
1591
1592 static int trace__tool_process(struct perf_tool *tool,
1593                                union perf_event *event,
1594                                struct perf_sample *sample,
1595                                struct machine *machine)
1596 {
1597         struct trace *trace = container_of(tool, struct trace, tool);
1598         return trace__process_event(trace, machine, event, sample);
1599 }
1600
1601 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1602 {
1603         struct machine *machine = vmachine;
1604
1605         if (machine->kptr_restrict_warned)
1606                 return NULL;
1607
1608         if (symbol_conf.kptr_restrict) {
1609                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1610                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1611                            "Kernel samples will not be resolved.\n");
1612                 machine->kptr_restrict_warned = true;
1613                 return NULL;
1614         }
1615
1616         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1617 }
1618
1619 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1620 {
1621         int err = symbol__init(NULL);
1622
1623         if (err)
1624                 return err;
1625
1626         trace->host = machine__new_host();
1627         if (trace->host == NULL)
1628                 return -ENOMEM;
1629
1630         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1631         if (err < 0)
1632                 goto out;
1633
1634         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1635                                             evlist->core.threads, trace__tool_process,
1636                                             true, false, 1);
1637 out:
1638         if (err)
1639                 symbol__exit();
1640
1641         return err;
1642 }
1643
1644 static void trace__symbols__exit(struct trace *trace)
1645 {
1646         machine__exit(trace->host);
1647         trace->host = NULL;
1648
1649         symbol__exit();
1650 }
1651
1652 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1653 {
1654         int idx;
1655
1656         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1657                 nr_args = sc->fmt->nr_args;
1658
1659         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1660         if (sc->arg_fmt == NULL)
1661                 return -1;
1662
1663         for (idx = 0; idx < nr_args; ++idx) {
1664                 if (sc->fmt)
1665                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1666         }
1667
1668         sc->nr_args = nr_args;
1669         return 0;
1670 }
1671
1672 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1673         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1674         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1675 };
1676
1677 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1678 {
1679        const struct syscall_arg_fmt *fmt = fmtp;
1680        return strcmp(name, fmt->name);
1681 }
1682
1683 static struct syscall_arg_fmt *
1684 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1685 {
1686        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1687 }
1688
1689 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1690 {
1691        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1692        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1693 }
1694
1695 static struct tep_format_field *
1696 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1697 {
1698         struct tep_format_field *last_field = NULL;
1699         int len;
1700
1701         for (; field; field = field->next, ++arg) {
1702                 last_field = field;
1703
1704                 if (arg->scnprintf)
1705                         continue;
1706
1707                 len = strlen(field->name);
1708
1709                 if (strcmp(field->type, "const char *") == 0 &&
1710                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1711                      strstr(field->name, "path") != NULL))
1712                         arg->scnprintf = SCA_FILENAME;
1713                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1714                         arg->scnprintf = SCA_PTR;
1715                 else if (strcmp(field->type, "pid_t") == 0)
1716                         arg->scnprintf = SCA_PID;
1717                 else if (strcmp(field->type, "umode_t") == 0)
1718                         arg->scnprintf = SCA_MODE_T;
1719                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1720                         arg->scnprintf = SCA_CHAR_ARRAY;
1721                         arg->nr_entries = field->arraylen;
1722                 } else if ((strcmp(field->type, "int") == 0 ||
1723                           strcmp(field->type, "unsigned int") == 0 ||
1724                           strcmp(field->type, "long") == 0) &&
1725                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1726                         /*
1727                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1728                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1729                          * 65 int
1730                          * 23 unsigned int
1731                          * 7 unsigned long
1732                          */
1733                         arg->scnprintf = SCA_FD;
1734                } else {
1735                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1736
1737                         if (fmt) {
1738                                 arg->scnprintf = fmt->scnprintf;
1739                                 arg->strtoul   = fmt->strtoul;
1740                         }
1741                 }
1742         }
1743
1744         return last_field;
1745 }
1746
1747 static int syscall__set_arg_fmts(struct syscall *sc)
1748 {
1749         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1750
1751         if (last_field)
1752                 sc->args_size = last_field->offset + last_field->size;
1753
1754         return 0;
1755 }
1756
1757 static int trace__read_syscall_info(struct trace *trace, int id)
1758 {
1759         char tp_name[128];
1760         struct syscall *sc;
1761         const char *name = syscalltbl__name(trace->sctbl, id);
1762
1763 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1764         if (trace->syscalls.table == NULL) {
1765                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1766                 if (trace->syscalls.table == NULL)
1767                         return -ENOMEM;
1768         }
1769 #else
1770         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1771                 // When using libaudit we don't know beforehand what is the max syscall id
1772                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1773
1774                 if (table == NULL)
1775                         return -ENOMEM;
1776
1777                 // Need to memset from offset 0 and +1 members if brand new
1778                 if (trace->syscalls.table == NULL)
1779                         memset(table, 0, (id + 1) * sizeof(*sc));
1780                 else
1781                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1782
1783                 trace->syscalls.table         = table;
1784                 trace->sctbl->syscalls.max_id = id;
1785         }
1786 #endif
1787         sc = trace->syscalls.table + id;
1788         if (sc->nonexistent)
1789                 return 0;
1790
1791         if (name == NULL) {
1792                 sc->nonexistent = true;
1793                 return 0;
1794         }
1795
1796         sc->name = name;
1797         sc->fmt  = syscall_fmt__find(sc->name);
1798
1799         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1800         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1801
1802         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1803                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1804                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1805         }
1806
1807         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1808                 return -ENOMEM;
1809
1810         if (IS_ERR(sc->tp_format))
1811                 return PTR_ERR(sc->tp_format);
1812
1813         sc->args = sc->tp_format->format.fields;
1814         /*
1815          * We need to check and discard the first variable '__syscall_nr'
1816          * or 'nr' that mean the syscall number. It is needless here.
1817          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1818          */
1819         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1820                 sc->args = sc->args->next;
1821                 --sc->nr_args;
1822         }
1823
1824         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1825         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1826
1827         return syscall__set_arg_fmts(sc);
1828 }
1829
1830 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1831 {
1832         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1833
1834         if (fmt != NULL) {
1835                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1836                 return 0;
1837         }
1838
1839         return -ENOMEM;
1840 }
1841
1842 static int intcmp(const void *a, const void *b)
1843 {
1844         const int *one = a, *another = b;
1845
1846         return *one - *another;
1847 }
1848
1849 static int trace__validate_ev_qualifier(struct trace *trace)
1850 {
1851         int err = 0;
1852         bool printed_invalid_prefix = false;
1853         struct str_node *pos;
1854         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1855
1856         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1857                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1858
1859         if (trace->ev_qualifier_ids.entries == NULL) {
1860                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1861                        trace->output);
1862                 err = -EINVAL;
1863                 goto out;
1864         }
1865
1866         strlist__for_each_entry(pos, trace->ev_qualifier) {
1867                 const char *sc = pos->s;
1868                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1869
1870                 if (id < 0) {
1871                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1872                         if (id >= 0)
1873                                 goto matches;
1874
1875                         if (!printed_invalid_prefix) {
1876                                 pr_debug("Skipping unknown syscalls: ");
1877                                 printed_invalid_prefix = true;
1878                         } else {
1879                                 pr_debug(", ");
1880                         }
1881
1882                         pr_debug("%s", sc);
1883                         continue;
1884                 }
1885 matches:
1886                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1887                 if (match_next == -1)
1888                         continue;
1889
1890                 while (1) {
1891                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1892                         if (id < 0)
1893                                 break;
1894                         if (nr_allocated == nr_used) {
1895                                 void *entries;
1896
1897                                 nr_allocated += 8;
1898                                 entries = realloc(trace->ev_qualifier_ids.entries,
1899                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1900                                 if (entries == NULL) {
1901                                         err = -ENOMEM;
1902                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1903                                         goto out_free;
1904                                 }
1905                                 trace->ev_qualifier_ids.entries = entries;
1906                         }
1907                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1908                 }
1909         }
1910
1911         trace->ev_qualifier_ids.nr = nr_used;
1912         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1913 out:
1914         if (printed_invalid_prefix)
1915                 pr_debug("\n");
1916         return err;
1917 out_free:
1918         zfree(&trace->ev_qualifier_ids.entries);
1919         trace->ev_qualifier_ids.nr = 0;
1920         goto out;
1921 }
1922
1923 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1924 {
1925         bool in_ev_qualifier;
1926
1927         if (trace->ev_qualifier_ids.nr == 0)
1928                 return true;
1929
1930         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1931                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1932
1933         if (in_ev_qualifier)
1934                return !trace->not_ev_qualifier;
1935
1936         return trace->not_ev_qualifier;
1937 }
1938
1939 /*
1940  * args is to be interpreted as a series of longs but we need to handle
1941  * 8-byte unaligned accesses. args points to raw_data within the event
1942  * and raw_data is guaranteed to be 8-byte unaligned because it is
1943  * preceded by raw_size which is a u32. So we need to copy args to a temp
1944  * variable to read it. Most notably this avoids extended load instructions
1945  * on unaligned addresses
1946  */
1947 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1948 {
1949         unsigned long val;
1950         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1951
1952         memcpy(&val, p, sizeof(val));
1953         return val;
1954 }
1955
1956 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1957                                       struct syscall_arg *arg)
1958 {
1959         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1960                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1961
1962         return scnprintf(bf, size, "arg%d: ", arg->idx);
1963 }
1964
1965 /*
1966  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1967  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1968  * in tools/perf/trace/beauty/mount_flags.c
1969  */
1970 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1971 {
1972         if (fmt && fmt->mask_val)
1973                 return fmt->mask_val(arg, val);
1974
1975         return val;
1976 }
1977
1978 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1979                                              struct syscall_arg *arg, unsigned long val)
1980 {
1981         if (fmt && fmt->scnprintf) {
1982                 arg->val = val;
1983                 if (fmt->parm)
1984                         arg->parm = fmt->parm;
1985                 return fmt->scnprintf(bf, size, arg);
1986         }
1987         return scnprintf(bf, size, "%ld", val);
1988 }
1989
1990 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1991                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1992                                       struct trace *trace, struct thread *thread)
1993 {
1994         size_t printed = 0;
1995         unsigned long val;
1996         u8 bit = 1;
1997         struct syscall_arg arg = {
1998                 .args   = args,
1999                 .augmented = {
2000                         .size = augmented_args_size,
2001                         .args = augmented_args,
2002                 },
2003                 .idx    = 0,
2004                 .mask   = 0,
2005                 .trace  = trace,
2006                 .thread = thread,
2007                 .show_string_prefix = trace->show_string_prefix,
2008         };
2009         struct thread_trace *ttrace = thread__priv(thread);
2010
2011         /*
2012          * Things like fcntl will set this in its 'cmd' formatter to pick the
2013          * right formatter for the return value (an fd? file flags?), which is
2014          * not needed for syscalls that always return a given type, say an fd.
2015          */
2016         ttrace->ret_scnprintf = NULL;
2017
2018         if (sc->args != NULL) {
2019                 struct tep_format_field *field;
2020
2021                 for (field = sc->args; field;
2022                      field = field->next, ++arg.idx, bit <<= 1) {
2023                         if (arg.mask & bit)
2024                                 continue;
2025
2026                         arg.fmt = &sc->arg_fmt[arg.idx];
2027                         val = syscall_arg__val(&arg, arg.idx);
2028                         /*
2029                          * Some syscall args need some mask, most don't and
2030                          * return val untouched.
2031                          */
2032                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2033
2034                         /*
2035                          * Suppress this argument if its value is zero and
2036                          * and we don't have a string associated in an
2037                          * strarray for it.
2038                          */
2039                         if (val == 0 &&
2040                             !trace->show_zeros &&
2041                             !(sc->arg_fmt &&
2042                               (sc->arg_fmt[arg.idx].show_zero ||
2043                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2044                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2045                               sc->arg_fmt[arg.idx].parm))
2046                                 continue;
2047
2048                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2049
2050                         if (trace->show_arg_names)
2051                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2052
2053                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2054                                                                   bf + printed, size - printed, &arg, val);
2055                 }
2056         } else if (IS_ERR(sc->tp_format)) {
2057                 /*
2058                  * If we managed to read the tracepoint /format file, then we
2059                  * may end up not having any args, like with gettid(), so only
2060                  * print the raw args when we didn't manage to read it.
2061                  */
2062                 while (arg.idx < sc->nr_args) {
2063                         if (arg.mask & bit)
2064                                 goto next_arg;
2065                         val = syscall_arg__val(&arg, arg.idx);
2066                         if (printed)
2067                                 printed += scnprintf(bf + printed, size - printed, ", ");
2068                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2069                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2070 next_arg:
2071                         ++arg.idx;
2072                         bit <<= 1;
2073                 }
2074         }
2075
2076         return printed;
2077 }
2078
2079 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2080                                   union perf_event *event,
2081                                   struct perf_sample *sample);
2082
2083 static struct syscall *trace__syscall_info(struct trace *trace,
2084                                            struct evsel *evsel, int id)
2085 {
2086         int err = 0;
2087
2088         if (id < 0) {
2089
2090                 /*
2091                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2092                  * before that, leaving at a higher verbosity level till that is
2093                  * explained. Reproduced with plain ftrace with:
2094                  *
2095                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2096                  * grep "NR -1 " /t/trace_pipe
2097                  *
2098                  * After generating some load on the machine.
2099                  */
2100                 if (verbose > 1) {
2101                         static u64 n;
2102                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2103                                 id, evsel__name(evsel), ++n);
2104                 }
2105                 return NULL;
2106         }
2107
2108         err = -EINVAL;
2109
2110 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2111         if (id > trace->sctbl->syscalls.max_id) {
2112 #else
2113         if (id >= trace->sctbl->syscalls.max_id) {
2114                 /*
2115                  * With libaudit we don't know beforehand what is the max_id,
2116                  * so we let trace__read_syscall_info() figure that out as we
2117                  * go on reading syscalls.
2118                  */
2119                 err = trace__read_syscall_info(trace, id);
2120                 if (err)
2121 #endif
2122                 goto out_cant_read;
2123         }
2124
2125         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2126             (err = trace__read_syscall_info(trace, id)) != 0)
2127                 goto out_cant_read;
2128
2129         if (trace->syscalls.table[id].name == NULL) {
2130                 if (trace->syscalls.table[id].nonexistent)
2131                         return NULL;
2132                 goto out_cant_read;
2133         }
2134
2135         return &trace->syscalls.table[id];
2136
2137 out_cant_read:
2138         if (verbose > 0) {
2139                 char sbuf[STRERR_BUFSIZE];
2140                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2141                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2142                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2143                 fputs(" information\n", trace->output);
2144         }
2145         return NULL;
2146 }
2147
2148 struct syscall_stats {
2149         struct stats stats;
2150         u64          nr_failures;
2151         int          max_errno;
2152         u32          *errnos;
2153 };
2154
2155 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2156                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2157 {
2158         struct int_node *inode;
2159         struct syscall_stats *stats;
2160         u64 duration = 0;
2161
2162         inode = intlist__findnew(ttrace->syscall_stats, id);
2163         if (inode == NULL)
2164                 return;
2165
2166         stats = inode->priv;
2167         if (stats == NULL) {
2168                 stats = malloc(sizeof(*stats));
2169                 if (stats == NULL)
2170                         return;
2171
2172                 stats->nr_failures = 0;
2173                 stats->max_errno   = 0;
2174                 stats->errnos      = NULL;
2175                 init_stats(&stats->stats);
2176                 inode->priv = stats;
2177         }
2178
2179         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2180                 duration = sample->time - ttrace->entry_time;
2181
2182         update_stats(&stats->stats, duration);
2183
2184         if (err < 0) {
2185                 ++stats->nr_failures;
2186
2187                 if (!errno_summary)
2188                         return;
2189
2190                 err = -err;
2191                 if (err > stats->max_errno) {
2192                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2193
2194                         if (new_errnos) {
2195                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2196                         } else {
2197                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2198                                          thread__comm_str(thread), thread->pid_, thread->tid);
2199                                 return;
2200                         }
2201
2202                         stats->errnos = new_errnos;
2203                         stats->max_errno = err;
2204                 }
2205
2206                 ++stats->errnos[err - 1];
2207         }
2208 }
2209
2210 static int trace__printf_interrupted_entry(struct trace *trace)
2211 {
2212         struct thread_trace *ttrace;
2213         size_t printed;
2214         int len;
2215
2216         if (trace->failure_only || trace->current == NULL)
2217                 return 0;
2218
2219         ttrace = thread__priv(trace->current);
2220
2221         if (!ttrace->entry_pending)
2222                 return 0;
2223
2224         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2225         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2226
2227         if (len < trace->args_alignment - 4)
2228                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2229
2230         printed += fprintf(trace->output, " ...\n");
2231
2232         ttrace->entry_pending = false;
2233         ++trace->nr_events_printed;
2234
2235         return printed;
2236 }
2237
2238 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2239                                  struct perf_sample *sample, struct thread *thread)
2240 {
2241         int printed = 0;
2242
2243         if (trace->print_sample) {
2244                 double ts = (double)sample->time / NSEC_PER_MSEC;
2245
2246                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2247                                    evsel__name(evsel), ts,
2248                                    thread__comm_str(thread),
2249                                    sample->pid, sample->tid, sample->cpu);
2250         }
2251
2252         return printed;
2253 }
2254
2255 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2256 {
2257         void *augmented_args = NULL;
2258         /*
2259          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2260          * and there we get all 6 syscall args plus the tracepoint common fields
2261          * that gets calculated at the start and the syscall_nr (another long).
2262          * So we check if that is the case and if so don't look after the
2263          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2264          * which is fixed.
2265          *
2266          * We'll revisit this later to pass s->args_size to the BPF augmenter
2267          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2268          * copies only what we need for each syscall, like what happens when we
2269          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2270          * traffic to just what is needed for each syscall.
2271          */
2272         int args_size = raw_augmented_args_size ?: sc->args_size;
2273
2274         *augmented_args_size = sample->raw_size - args_size;
2275         if (*augmented_args_size > 0)
2276                 augmented_args = sample->raw_data + args_size;
2277
2278         return augmented_args;
2279 }
2280
2281 static void syscall__exit(struct syscall *sc)
2282 {
2283         if (!sc)
2284                 return;
2285
2286         free(sc->arg_fmt);
2287 }
2288
2289 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2290                             union perf_event *event __maybe_unused,
2291                             struct perf_sample *sample)
2292 {
2293         char *msg;
2294         void *args;
2295         int printed = 0;
2296         struct thread *thread;
2297         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2298         int augmented_args_size = 0;
2299         void *augmented_args = NULL;
2300         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2301         struct thread_trace *ttrace;
2302
2303         if (sc == NULL)
2304                 return -1;
2305
2306         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2307         ttrace = thread__trace(thread, trace->output);
2308         if (ttrace == NULL)
2309                 goto out_put;
2310
2311         trace__fprintf_sample(trace, evsel, sample, thread);
2312
2313         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2314
2315         if (ttrace->entry_str == NULL) {
2316                 ttrace->entry_str = malloc(trace__entry_str_size);
2317                 if (!ttrace->entry_str)
2318                         goto out_put;
2319         }
2320
2321         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2322                 trace__printf_interrupted_entry(trace);
2323         /*
2324          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2325          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2326          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2327          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2328          * so when handling, say the openat syscall, we end up getting 6 args for the
2329          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2330          * thinking that the extra 2 u64 args are the augmented filename, so just check
2331          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2332          */
2333         if (evsel != trace->syscalls.events.sys_enter)
2334                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2335         ttrace->entry_time = sample->time;
2336         msg = ttrace->entry_str;
2337         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2338
2339         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2340                                            args, augmented_args, augmented_args_size, trace, thread);
2341
2342         if (sc->is_exit) {
2343                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2344                         int alignment = 0;
2345
2346                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2347                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2348                         if (trace->args_alignment > printed)
2349                                 alignment = trace->args_alignment - printed;
2350                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2351                 }
2352         } else {
2353                 ttrace->entry_pending = true;
2354                 /* See trace__vfs_getname & trace__sys_exit */
2355                 ttrace->filename.pending_open = false;
2356         }
2357
2358         if (trace->current != thread) {
2359                 thread__put(trace->current);
2360                 trace->current = thread__get(thread);
2361         }
2362         err = 0;
2363 out_put:
2364         thread__put(thread);
2365         return err;
2366 }
2367
2368 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2369                                     struct perf_sample *sample)
2370 {
2371         struct thread_trace *ttrace;
2372         struct thread *thread;
2373         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2374         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2375         char msg[1024];
2376         void *args, *augmented_args = NULL;
2377         int augmented_args_size;
2378
2379         if (sc == NULL)
2380                 return -1;
2381
2382         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2383         ttrace = thread__trace(thread, trace->output);
2384         /*
2385          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2386          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2387          */
2388         if (ttrace == NULL)
2389                 goto out_put;
2390
2391         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2392         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2393         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2394         fprintf(trace->output, "%s", msg);
2395         err = 0;
2396 out_put:
2397         thread__put(thread);
2398         return err;
2399 }
2400
2401 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2402                                     struct perf_sample *sample,
2403                                     struct callchain_cursor *cursor)
2404 {
2405         struct addr_location al;
2406         int max_stack = evsel->core.attr.sample_max_stack ?
2407                         evsel->core.attr.sample_max_stack :
2408                         trace->max_stack;
2409         int err;
2410
2411         if (machine__resolve(trace->host, &al, sample) < 0)
2412                 return -1;
2413
2414         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2415         addr_location__put(&al);
2416         return err;
2417 }
2418
2419 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2420 {
2421         /* TODO: user-configurable print_opts */
2422         const unsigned int print_opts = EVSEL__PRINT_SYM |
2423                                         EVSEL__PRINT_DSO |
2424                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2425
2426         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2427 }
2428
2429 static const char *errno_to_name(struct evsel *evsel, int err)
2430 {
2431         struct perf_env *env = evsel__env(evsel);
2432         const char *arch_name = perf_env__arch(env);
2433
2434         return arch_syscalls__strerrno(arch_name, err);
2435 }
2436
2437 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2438                            union perf_event *event __maybe_unused,
2439                            struct perf_sample *sample)
2440 {
2441         long ret;
2442         u64 duration = 0;
2443         bool duration_calculated = false;
2444         struct thread *thread;
2445         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2446         int alignment = trace->args_alignment;
2447         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2448         struct thread_trace *ttrace;
2449
2450         if (sc == NULL)
2451                 return -1;
2452
2453         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2454         ttrace = thread__trace(thread, trace->output);
2455         if (ttrace == NULL)
2456                 goto out_put;
2457
2458         trace__fprintf_sample(trace, evsel, sample, thread);
2459
2460         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2461
2462         if (trace->summary)
2463                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2464
2465         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2466                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2467                 ttrace->filename.pending_open = false;
2468                 ++trace->stats.vfs_getname;
2469         }
2470
2471         if (ttrace->entry_time) {
2472                 duration = sample->time - ttrace->entry_time;
2473                 if (trace__filter_duration(trace, duration))
2474                         goto out;
2475                 duration_calculated = true;
2476         } else if (trace->duration_filter)
2477                 goto out;
2478
2479         if (sample->callchain) {
2480                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2481                 if (callchain_ret == 0) {
2482                         if (callchain_cursor.nr < trace->min_stack)
2483                                 goto out;
2484                         callchain_ret = 1;
2485                 }
2486         }
2487
2488         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2489                 goto out;
2490
2491         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2492
2493         if (ttrace->entry_pending) {
2494                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2495         } else {
2496                 printed += fprintf(trace->output, " ... [");
2497                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2498                 printed += 9;
2499                 printed += fprintf(trace->output, "]: %s()", sc->name);
2500         }
2501
2502         printed++; /* the closing ')' */
2503
2504         if (alignment > printed)
2505                 alignment -= printed;
2506         else
2507                 alignment = 0;
2508
2509         fprintf(trace->output, ")%*s= ", alignment, " ");
2510
2511         if (sc->fmt == NULL) {
2512                 if (ret < 0)
2513                         goto errno_print;
2514 signed_print:
2515                 fprintf(trace->output, "%ld", ret);
2516         } else if (ret < 0) {
2517 errno_print: {
2518                 char bf[STRERR_BUFSIZE];
2519                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2520                            *e = errno_to_name(evsel, -ret);
2521
2522                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2523         }
2524         } else if (ret == 0 && sc->fmt->timeout)
2525                 fprintf(trace->output, "0 (Timeout)");
2526         else if (ttrace->ret_scnprintf) {
2527                 char bf[1024];
2528                 struct syscall_arg arg = {
2529                         .val    = ret,
2530                         .thread = thread,
2531                         .trace  = trace,
2532                 };
2533                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2534                 ttrace->ret_scnprintf = NULL;
2535                 fprintf(trace->output, "%s", bf);
2536         } else if (sc->fmt->hexret)
2537                 fprintf(trace->output, "%#lx", ret);
2538         else if (sc->fmt->errpid) {
2539                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2540
2541                 if (child != NULL) {
2542                         fprintf(trace->output, "%ld", ret);
2543                         if (child->comm_set)
2544                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2545                         thread__put(child);
2546                 }
2547         } else
2548                 goto signed_print;
2549
2550         fputc('\n', trace->output);
2551
2552         /*
2553          * We only consider an 'event' for the sake of --max-events a non-filtered
2554          * sys_enter + sys_exit and other tracepoint events.
2555          */
2556         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2557                 interrupted = true;
2558
2559         if (callchain_ret > 0)
2560                 trace__fprintf_callchain(trace, sample);
2561         else if (callchain_ret < 0)
2562                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2563 out:
2564         ttrace->entry_pending = false;
2565         err = 0;
2566 out_put:
2567         thread__put(thread);
2568         return err;
2569 }
2570
2571 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2572                               union perf_event *event __maybe_unused,
2573                               struct perf_sample *sample)
2574 {
2575         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2576         struct thread_trace *ttrace;
2577         size_t filename_len, entry_str_len, to_move;
2578         ssize_t remaining_space;
2579         char *pos;
2580         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2581
2582         if (!thread)
2583                 goto out;
2584
2585         ttrace = thread__priv(thread);
2586         if (!ttrace)
2587                 goto out_put;
2588
2589         filename_len = strlen(filename);
2590         if (filename_len == 0)
2591                 goto out_put;
2592
2593         if (ttrace->filename.namelen < filename_len) {
2594                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2595
2596                 if (f == NULL)
2597                         goto out_put;
2598
2599                 ttrace->filename.namelen = filename_len;
2600                 ttrace->filename.name = f;
2601         }
2602
2603         strcpy(ttrace->filename.name, filename);
2604         ttrace->filename.pending_open = true;
2605
2606         if (!ttrace->filename.ptr)
2607                 goto out_put;
2608
2609         entry_str_len = strlen(ttrace->entry_str);
2610         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2611         if (remaining_space <= 0)
2612                 goto out_put;
2613
2614         if (filename_len > (size_t)remaining_space) {
2615                 filename += filename_len - remaining_space;
2616                 filename_len = remaining_space;
2617         }
2618
2619         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2620         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2621         memmove(pos + filename_len, pos, to_move);
2622         memcpy(pos, filename, filename_len);
2623
2624         ttrace->filename.ptr = 0;
2625         ttrace->filename.entry_str_pos = 0;
2626 out_put:
2627         thread__put(thread);
2628 out:
2629         return 0;
2630 }
2631
2632 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2633                                      union perf_event *event __maybe_unused,
2634                                      struct perf_sample *sample)
2635 {
2636         u64 runtime = evsel__intval(evsel, sample, "runtime");
2637         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2638         struct thread *thread = machine__findnew_thread(trace->host,
2639                                                         sample->pid,
2640                                                         sample->tid);
2641         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2642
2643         if (ttrace == NULL)
2644                 goto out_dump;
2645
2646         ttrace->runtime_ms += runtime_ms;
2647         trace->runtime_ms += runtime_ms;
2648 out_put:
2649         thread__put(thread);
2650         return 0;
2651
2652 out_dump:
2653         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2654                evsel->name,
2655                evsel__strval(evsel, sample, "comm"),
2656                (pid_t)evsel__intval(evsel, sample, "pid"),
2657                runtime,
2658                evsel__intval(evsel, sample, "vruntime"));
2659         goto out_put;
2660 }
2661
2662 static int bpf_output__printer(enum binary_printer_ops op,
2663                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2664 {
2665         unsigned char ch = (unsigned char)val;
2666
2667         switch (op) {
2668         case BINARY_PRINT_CHAR_DATA:
2669                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2670         case BINARY_PRINT_DATA_BEGIN:
2671         case BINARY_PRINT_LINE_BEGIN:
2672         case BINARY_PRINT_ADDR:
2673         case BINARY_PRINT_NUM_DATA:
2674         case BINARY_PRINT_NUM_PAD:
2675         case BINARY_PRINT_SEP:
2676         case BINARY_PRINT_CHAR_PAD:
2677         case BINARY_PRINT_LINE_END:
2678         case BINARY_PRINT_DATA_END:
2679         default:
2680                 break;
2681         }
2682
2683         return 0;
2684 }
2685
2686 static void bpf_output__fprintf(struct trace *trace,
2687                                 struct perf_sample *sample)
2688 {
2689         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2690                         bpf_output__printer, NULL, trace->output);
2691         ++trace->nr_events_printed;
2692 }
2693
2694 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2695                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2696 {
2697         char bf[2048];
2698         size_t size = sizeof(bf);
2699         struct tep_format_field *field = evsel->tp_format->format.fields;
2700         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2701         size_t printed = 0;
2702         unsigned long val;
2703         u8 bit = 1;
2704         struct syscall_arg syscall_arg = {
2705                 .augmented = {
2706                         .size = augmented_args_size,
2707                         .args = augmented_args,
2708                 },
2709                 .idx    = 0,
2710                 .mask   = 0,
2711                 .trace  = trace,
2712                 .thread = thread,
2713                 .show_string_prefix = trace->show_string_prefix,
2714         };
2715
2716         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2717                 if (syscall_arg.mask & bit)
2718                         continue;
2719
2720                 syscall_arg.len = 0;
2721                 syscall_arg.fmt = arg;
2722                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2723                         int offset = field->offset;
2724
2725                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2726                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2727                                 syscall_arg.len = offset >> 16;
2728                                 offset &= 0xffff;
2729                         }
2730
2731                         val = (uintptr_t)(sample->raw_data + offset);
2732                 } else
2733                         val = format_field__intval(field, sample, evsel->needs_swap);
2734                 /*
2735                  * Some syscall args need some mask, most don't and
2736                  * return val untouched.
2737                  */
2738                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2739
2740                 /*
2741                  * Suppress this argument if its value is zero and
2742                  * and we don't have a string associated in an
2743                  * strarray for it.
2744                  */
2745                 if (val == 0 &&
2746                     !trace->show_zeros &&
2747                     !((arg->show_zero ||
2748                        arg->scnprintf == SCA_STRARRAY ||
2749                        arg->scnprintf == SCA_STRARRAYS) &&
2750                       arg->parm))
2751                         continue;
2752
2753                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2754
2755                 /*
2756                  * XXX Perhaps we should have a show_tp_arg_names,
2757                  * leaving show_arg_names just for syscalls?
2758                  */
2759                 if (1 || trace->show_arg_names)
2760                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2761
2762                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2763         }
2764
2765         return printed + fprintf(trace->output, "%s", bf);
2766 }
2767
2768 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2769                                 union perf_event *event __maybe_unused,
2770                                 struct perf_sample *sample)
2771 {
2772         struct thread *thread;
2773         int callchain_ret = 0;
2774         /*
2775          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2776          * this event's max_events having been hit and this is an entry coming
2777          * from the ring buffer that we should discard, since the max events
2778          * have already been considered/printed.
2779          */
2780         if (evsel->disabled)
2781                 return 0;
2782
2783         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2784
2785         if (sample->callchain) {
2786                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2787                 if (callchain_ret == 0) {
2788                         if (callchain_cursor.nr < trace->min_stack)
2789                                 goto out;
2790                         callchain_ret = 1;
2791                 }
2792         }
2793
2794         trace__printf_interrupted_entry(trace);
2795         trace__fprintf_tstamp(trace, sample->time, trace->output);
2796
2797         if (trace->trace_syscalls && trace->show_duration)
2798                 fprintf(trace->output, "(         ): ");
2799
2800         if (thread)
2801                 trace__fprintf_comm_tid(trace, thread, trace->output);
2802
2803         if (evsel == trace->syscalls.events.augmented) {
2804                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2805                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2806
2807                 if (sc) {
2808                         fprintf(trace->output, "%s(", sc->name);
2809                         trace__fprintf_sys_enter(trace, evsel, sample);
2810                         fputc(')', trace->output);
2811                         goto newline;
2812                 }
2813
2814                 /*
2815                  * XXX: Not having the associated syscall info or not finding/adding
2816                  *      the thread should never happen, but if it does...
2817                  *      fall thru and print it as a bpf_output event.
2818                  */
2819         }
2820
2821         fprintf(trace->output, "%s(", evsel->name);
2822
2823         if (evsel__is_bpf_output(evsel)) {
2824                 bpf_output__fprintf(trace, sample);
2825         } else if (evsel->tp_format) {
2826                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2827                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2828                         if (trace->libtraceevent_print) {
2829                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2830                                                       sample->raw_data, sample->raw_size,
2831                                                       trace->output);
2832                         } else {
2833                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2834                         }
2835                 }
2836         }
2837
2838 newline:
2839         fprintf(trace->output, ")\n");
2840
2841         if (callchain_ret > 0)
2842                 trace__fprintf_callchain(trace, sample);
2843         else if (callchain_ret < 0)
2844                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2845
2846         ++trace->nr_events_printed;
2847
2848         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2849                 evsel__disable(evsel);
2850                 evsel__close(evsel);
2851         }
2852 out:
2853         thread__put(thread);
2854         return 0;
2855 }
2856
2857 static void print_location(FILE *f, struct perf_sample *sample,
2858                            struct addr_location *al,
2859                            bool print_dso, bool print_sym)
2860 {
2861
2862         if ((verbose > 0 || print_dso) && al->map)
2863                 fprintf(f, "%s@", al->map->dso->long_name);
2864
2865         if ((verbose > 0 || print_sym) && al->sym)
2866                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2867                         al->addr - al->sym->start);
2868         else if (al->map)
2869                 fprintf(f, "0x%" PRIx64, al->addr);
2870         else
2871                 fprintf(f, "0x%" PRIx64, sample->addr);
2872 }
2873
2874 static int trace__pgfault(struct trace *trace,
2875                           struct evsel *evsel,
2876                           union perf_event *event __maybe_unused,
2877                           struct perf_sample *sample)
2878 {
2879         struct thread *thread;
2880         struct addr_location al;
2881         char map_type = 'd';
2882         struct thread_trace *ttrace;
2883         int err = -1;
2884         int callchain_ret = 0;
2885
2886         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2887
2888         if (sample->callchain) {
2889                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2890                 if (callchain_ret == 0) {
2891                         if (callchain_cursor.nr < trace->min_stack)
2892                                 goto out_put;
2893                         callchain_ret = 1;
2894                 }
2895         }
2896
2897         ttrace = thread__trace(thread, trace->output);
2898         if (ttrace == NULL)
2899                 goto out_put;
2900
2901         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2902                 ttrace->pfmaj++;
2903         else
2904                 ttrace->pfmin++;
2905
2906         if (trace->summary_only)
2907                 goto out;
2908
2909         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2910
2911         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2912
2913         fprintf(trace->output, "%sfault [",
2914                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2915                 "maj" : "min");
2916
2917         print_location(trace->output, sample, &al, false, true);
2918
2919         fprintf(trace->output, "] => ");
2920
2921         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2922
2923         if (!al.map) {
2924                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2925
2926                 if (al.map)
2927                         map_type = 'x';
2928                 else
2929                         map_type = '?';
2930         }
2931
2932         print_location(trace->output, sample, &al, true, false);
2933
2934         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2935
2936         if (callchain_ret > 0)
2937                 trace__fprintf_callchain(trace, sample);
2938         else if (callchain_ret < 0)
2939                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2940
2941         ++trace->nr_events_printed;
2942 out:
2943         err = 0;
2944 out_put:
2945         thread__put(thread);
2946         return err;
2947 }
2948
2949 static void trace__set_base_time(struct trace *trace,
2950                                  struct evsel *evsel,
2951                                  struct perf_sample *sample)
2952 {
2953         /*
2954          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2955          * and don't use sample->time unconditionally, we may end up having
2956          * some other event in the future without PERF_SAMPLE_TIME for good
2957          * reason, i.e. we may not be interested in its timestamps, just in
2958          * it taking place, picking some piece of information when it
2959          * appears in our event stream (vfs_getname comes to mind).
2960          */
2961         if (trace->base_time == 0 && !trace->full_time &&
2962             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2963                 trace->base_time = sample->time;
2964 }
2965
2966 static int trace__process_sample(struct perf_tool *tool,
2967                                  union perf_event *event,
2968                                  struct perf_sample *sample,
2969                                  struct evsel *evsel,
2970                                  struct machine *machine __maybe_unused)
2971 {
2972         struct trace *trace = container_of(tool, struct trace, tool);
2973         struct thread *thread;
2974         int err = 0;
2975
2976         tracepoint_handler handler = evsel->handler;
2977
2978         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2979         if (thread && thread__is_filtered(thread))
2980                 goto out;
2981
2982         trace__set_base_time(trace, evsel, sample);
2983
2984         if (handler) {
2985                 ++trace->nr_events;
2986                 handler(trace, evsel, event, sample);
2987         }
2988 out:
2989         thread__put(thread);
2990         return err;
2991 }
2992
2993 static int trace__record(struct trace *trace, int argc, const char **argv)
2994 {
2995         unsigned int rec_argc, i, j;
2996         const char **rec_argv;
2997         const char * const record_args[] = {
2998                 "record",
2999                 "-R",
3000                 "-m", "1024",
3001                 "-c", "1",
3002         };
3003         pid_t pid = getpid();
3004         char *filter = asprintf__tp_filter_pids(1, &pid);
3005         const char * const sc_args[] = { "-e", };
3006         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3007         const char * const majpf_args[] = { "-e", "major-faults" };
3008         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3009         const char * const minpf_args[] = { "-e", "minor-faults" };
3010         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3011         int err = -1;
3012
3013         /* +3 is for the event string below and the pid filter */
3014         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3015                 majpf_args_nr + minpf_args_nr + argc;
3016         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3017
3018         if (rec_argv == NULL || filter == NULL)
3019                 goto out_free;
3020
3021         j = 0;
3022         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3023                 rec_argv[j++] = record_args[i];
3024
3025         if (trace->trace_syscalls) {
3026                 for (i = 0; i < sc_args_nr; i++)
3027                         rec_argv[j++] = sc_args[i];
3028
3029                 /* event string may be different for older kernels - e.g., RHEL6 */
3030                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3031                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3032                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3033                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3034                 else {
3035                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3036                         goto out_free;
3037                 }
3038         }
3039
3040         rec_argv[j++] = "--filter";
3041         rec_argv[j++] = filter;
3042
3043         if (trace->trace_pgfaults & TRACE_PFMAJ)
3044                 for (i = 0; i < majpf_args_nr; i++)
3045                         rec_argv[j++] = majpf_args[i];
3046
3047         if (trace->trace_pgfaults & TRACE_PFMIN)
3048                 for (i = 0; i < minpf_args_nr; i++)
3049                         rec_argv[j++] = minpf_args[i];
3050
3051         for (i = 0; i < (unsigned int)argc; i++)
3052                 rec_argv[j++] = argv[i];
3053
3054         err = cmd_record(j, rec_argv);
3055 out_free:
3056         free(filter);
3057         free(rec_argv);
3058         return err;
3059 }
3060
3061 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3062
3063 static bool evlist__add_vfs_getname(struct evlist *evlist)
3064 {
3065         bool found = false;
3066         struct evsel *evsel, *tmp;
3067         struct parse_events_error err;
3068         int ret;
3069
3070         parse_events_error__init(&err);
3071         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3072         parse_events_error__exit(&err);
3073         if (ret)
3074                 return false;
3075
3076         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3077                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3078                         continue;
3079
3080                 if (evsel__field(evsel, "pathname")) {
3081                         evsel->handler = trace__vfs_getname;
3082                         found = true;
3083                         continue;
3084                 }
3085
3086                 list_del_init(&evsel->core.node);
3087                 evsel->evlist = NULL;
3088                 evsel__delete(evsel);
3089         }
3090
3091         return found;
3092 }
3093
3094 static struct evsel *evsel__new_pgfault(u64 config)
3095 {
3096         struct evsel *evsel;
3097         struct perf_event_attr attr = {
3098                 .type = PERF_TYPE_SOFTWARE,
3099                 .mmap_data = 1,
3100         };
3101
3102         attr.config = config;
3103         attr.sample_period = 1;
3104
3105         event_attr_init(&attr);
3106
3107         evsel = evsel__new(&attr);
3108         if (evsel)
3109                 evsel->handler = trace__pgfault;
3110
3111         return evsel;
3112 }
3113
3114 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3115 {
3116         struct evsel *evsel;
3117
3118         evlist__for_each_entry(evlist, evsel) {
3119                 struct evsel_trace *et = evsel->priv;
3120
3121                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3122                         continue;
3123
3124                 free(et->fmt);
3125                 free(et);
3126         }
3127 }
3128
3129 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3130 {
3131         const u32 type = event->header.type;
3132         struct evsel *evsel;
3133
3134         if (type != PERF_RECORD_SAMPLE) {
3135                 trace__process_event(trace, trace->host, event, sample);
3136                 return;
3137         }
3138
3139         evsel = evlist__id2evsel(trace->evlist, sample->id);
3140         if (evsel == NULL) {
3141                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3142                 return;
3143         }
3144
3145         if (evswitch__discard(&trace->evswitch, evsel))
3146                 return;
3147
3148         trace__set_base_time(trace, evsel, sample);
3149
3150         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3151             sample->raw_data == NULL) {
3152                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3153                        evsel__name(evsel), sample->tid,
3154                        sample->cpu, sample->raw_size);
3155         } else {
3156                 tracepoint_handler handler = evsel->handler;
3157                 handler(trace, evsel, event, sample);
3158         }
3159
3160         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3161                 interrupted = true;
3162 }
3163
3164 static int trace__add_syscall_newtp(struct trace *trace)
3165 {
3166         int ret = -1;
3167         struct evlist *evlist = trace->evlist;
3168         struct evsel *sys_enter, *sys_exit;
3169
3170         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3171         if (sys_enter == NULL)
3172                 goto out;
3173
3174         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3175                 goto out_delete_sys_enter;
3176
3177         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3178         if (sys_exit == NULL)
3179                 goto out_delete_sys_enter;
3180
3181         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3182                 goto out_delete_sys_exit;
3183
3184         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3185         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3186
3187         evlist__add(evlist, sys_enter);
3188         evlist__add(evlist, sys_exit);
3189
3190         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3191                 /*
3192                  * We're interested only in the user space callchain
3193                  * leading to the syscall, allow overriding that for
3194                  * debugging reasons using --kernel_syscall_callchains
3195                  */
3196                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3197         }
3198
3199         trace->syscalls.events.sys_enter = sys_enter;
3200         trace->syscalls.events.sys_exit  = sys_exit;
3201
3202         ret = 0;
3203 out:
3204         return ret;
3205
3206 out_delete_sys_exit:
3207         evsel__delete_priv(sys_exit);
3208 out_delete_sys_enter:
3209         evsel__delete_priv(sys_enter);
3210         goto out;
3211 }
3212
3213 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3214 {
3215         int err = -1;
3216         struct evsel *sys_exit;
3217         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3218                                                 trace->ev_qualifier_ids.nr,
3219                                                 trace->ev_qualifier_ids.entries);
3220
3221         if (filter == NULL)
3222                 goto out_enomem;
3223
3224         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3225                 sys_exit = trace->syscalls.events.sys_exit;
3226                 err = evsel__append_tp_filter(sys_exit, filter);
3227         }
3228
3229         free(filter);
3230 out:
3231         return err;
3232 out_enomem:
3233         errno = ENOMEM;
3234         goto out;
3235 }
3236
3237 #ifdef HAVE_LIBBPF_SUPPORT
3238 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3239 {
3240         if (trace->bpf_obj == NULL)
3241                 return NULL;
3242
3243         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3244 }
3245
3246 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3247 {
3248         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3249 }
3250
3251 static void trace__set_bpf_map_syscalls(struct trace *trace)
3252 {
3253         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
3254         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3255         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3256 }
3257
3258 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3259 {
3260         struct bpf_program *pos, *prog = NULL;
3261         const char *sec_name;
3262
3263         if (trace->bpf_obj == NULL)
3264                 return NULL;
3265
3266         bpf_object__for_each_program(pos, trace->bpf_obj) {
3267                 sec_name = bpf_program__section_name(pos);
3268                 if (sec_name && !strcmp(sec_name, name)) {
3269                         prog = pos;
3270                         break;
3271                 }
3272         }
3273
3274         return prog;
3275 }
3276
3277 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3278                                                         const char *prog_name, const char *type)
3279 {
3280         struct bpf_program *prog;
3281
3282         if (prog_name == NULL) {
3283                 char default_prog_name[256];
3284                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3285                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3286                 if (prog != NULL)
3287                         goto out_found;
3288                 if (sc->fmt && sc->fmt->alias) {
3289                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3290                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3291                         if (prog != NULL)
3292                                 goto out_found;
3293                 }
3294                 goto out_unaugmented;
3295         }
3296
3297         prog = trace__find_bpf_program_by_title(trace, prog_name);
3298
3299         if (prog != NULL) {
3300 out_found:
3301                 return prog;
3302         }
3303
3304         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3305                  prog_name, type, sc->name);
3306 out_unaugmented:
3307         return trace->syscalls.unaugmented_prog;
3308 }
3309
3310 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3311 {
3312         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3313
3314         if (sc == NULL)
3315                 return;
3316
3317         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3318         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3319 }
3320
3321 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3322 {
3323         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3324         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3325 }
3326
3327 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3328 {
3329         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3330         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3331 }
3332
3333 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3334 {
3335         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3336         int arg = 0;
3337
3338         if (sc == NULL)
3339                 goto out;
3340
3341         for (; arg < sc->nr_args; ++arg) {
3342                 entry->string_args_len[arg] = 0;
3343                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3344                         /* Should be set like strace -s strsize */
3345                         entry->string_args_len[arg] = PATH_MAX;
3346                 }
3347         }
3348 out:
3349         for (; arg < 6; ++arg)
3350                 entry->string_args_len[arg] = 0;
3351 }
3352 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3353 {
3354         int fd = bpf_map__fd(trace->syscalls.map);
3355         struct bpf_map_syscall_entry value = {
3356                 .enabled = !trace->not_ev_qualifier,
3357         };
3358         int err = 0;
3359         size_t i;
3360
3361         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3362                 int key = trace->ev_qualifier_ids.entries[i];
3363
3364                 if (value.enabled) {
3365                         trace__init_bpf_map_syscall_args(trace, key, &value);
3366                         trace__init_syscall_bpf_progs(trace, key);
3367                 }
3368
3369                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3370                 if (err)
3371                         break;
3372         }
3373
3374         return err;
3375 }
3376
3377 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3378 {
3379         int fd = bpf_map__fd(trace->syscalls.map);
3380         struct bpf_map_syscall_entry value = {
3381                 .enabled = enabled,
3382         };
3383         int err = 0, key;
3384
3385         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3386                 if (enabled)
3387                         trace__init_bpf_map_syscall_args(trace, key, &value);
3388
3389                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3390                 if (err)
3391                         break;
3392         }
3393
3394         return err;
3395 }
3396
3397 static int trace__init_syscalls_bpf_map(struct trace *trace)
3398 {
3399         bool enabled = true;
3400
3401         if (trace->ev_qualifier_ids.nr)
3402                 enabled = trace->not_ev_qualifier;
3403
3404         return __trace__init_syscalls_bpf_map(trace, enabled);
3405 }
3406
3407 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3408 {
3409         struct tep_format_field *field, *candidate_field;
3410         int id;
3411
3412         /*
3413          * We're only interested in syscalls that have a pointer:
3414          */
3415         for (field = sc->args; field; field = field->next) {
3416                 if (field->flags & TEP_FIELD_IS_POINTER)
3417                         goto try_to_find_pair;
3418         }
3419
3420         return NULL;
3421
3422 try_to_find_pair:
3423         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3424                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3425                 struct bpf_program *pair_prog;
3426                 bool is_candidate = false;
3427
3428                 if (pair == NULL || pair == sc ||
3429                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3430                         continue;
3431
3432                 for (field = sc->args, candidate_field = pair->args;
3433                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3434                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3435                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3436
3437                         if (is_pointer) {
3438                                if (!candidate_is_pointer) {
3439                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3440                                         continue;
3441                                }
3442                         } else {
3443                                 if (candidate_is_pointer) {
3444                                         // The candidate might copy a pointer we don't have, skip it.
3445                                         goto next_candidate;
3446                                 }
3447                                 continue;
3448                         }
3449
3450                         if (strcmp(field->type, candidate_field->type))
3451                                 goto next_candidate;
3452
3453                         is_candidate = true;
3454                 }
3455
3456                 if (!is_candidate)
3457                         goto next_candidate;
3458
3459                 /*
3460                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3461                  * then it may be collecting that and we then can't use it, as it would collect
3462                  * more than what is common to the two syscalls.
3463                  */
3464                 if (candidate_field) {
3465                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3466                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3467                                         goto next_candidate;
3468                 }
3469
3470                 pair_prog = pair->bpf_prog.sys_enter;
3471                 /*
3472                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3473                  * have been searched for, so search it here and if it returns the
3474                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3475                  * program for a filtered syscall on a non-filtered one.
3476                  *
3477                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3478                  * useful for "renameat2".
3479                  */
3480                 if (pair_prog == NULL) {
3481                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3482                         if (pair_prog == trace->syscalls.unaugmented_prog)
3483                                 goto next_candidate;
3484                 }
3485
3486                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3487                 return pair_prog;
3488         next_candidate:
3489                 continue;
3490         }
3491
3492         return NULL;
3493 }
3494
3495 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3496 {
3497         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3498             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3499         int err = 0, key;
3500
3501         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3502                 int prog_fd;
3503
3504                 if (!trace__syscall_enabled(trace, key))
3505                         continue;
3506
3507                 trace__init_syscall_bpf_progs(trace, key);
3508
3509                 // It'll get at least the "!raw_syscalls:unaugmented"
3510                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3511                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3512                 if (err)
3513                         break;
3514                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3515                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3516                 if (err)
3517                         break;
3518         }
3519
3520         /*
3521          * Now lets do a second pass looking for enabled syscalls without
3522          * an augmenter that have a signature that is a superset of another
3523          * syscall with an augmenter so that we can auto-reuse it.
3524          *
3525          * I.e. if we have an augmenter for the "open" syscall that has
3526          * this signature:
3527          *
3528          *   int open(const char *pathname, int flags, mode_t mode);
3529          *
3530          * I.e. that will collect just the first string argument, then we
3531          * can reuse it for the 'creat' syscall, that has this signature:
3532          *
3533          *   int creat(const char *pathname, mode_t mode);
3534          *
3535          * and for:
3536          *
3537          *   int stat(const char *pathname, struct stat *statbuf);
3538          *   int lstat(const char *pathname, struct stat *statbuf);
3539          *
3540          * Because the 'open' augmenter will collect the first arg as a string,
3541          * and leave alone all the other args, which already helps with
3542          * beautifying 'stat' and 'lstat''s pathname arg.
3543          *
3544          * Then, in time, when 'stat' gets an augmenter that collects both
3545          * first and second arg (this one on the raw_syscalls:sys_exit prog
3546          * array tail call, then that one will be used.
3547          */
3548         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3549                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3550                 struct bpf_program *pair_prog;
3551                 int prog_fd;
3552
3553                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3554                         continue;
3555
3556                 /*
3557                  * For now we're just reusing the sys_enter prog, and if it
3558                  * already has an augmenter, we don't need to find one.
3559                  */
3560                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3561                         continue;
3562
3563                 /*
3564                  * Look at all the other syscalls for one that has a signature
3565                  * that is close enough that we can share:
3566                  */
3567                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3568                 if (pair_prog == NULL)
3569                         continue;
3570
3571                 sc->bpf_prog.sys_enter = pair_prog;
3572
3573                 /*
3574                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3575                  * with the fd for the program we're reusing:
3576                  */
3577                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3578                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3579                 if (err)
3580                         break;
3581         }
3582
3583
3584         return err;
3585 }
3586
3587 static void trace__delete_augmented_syscalls(struct trace *trace)
3588 {
3589         struct evsel *evsel, *tmp;
3590
3591         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3592         evsel__delete(trace->syscalls.events.augmented);
3593         trace->syscalls.events.augmented = NULL;
3594
3595         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3596                 if (evsel->bpf_obj == trace->bpf_obj) {
3597                         evlist__remove(trace->evlist, evsel);
3598                         evsel__delete(evsel);
3599                 }
3600
3601         }
3602
3603         bpf_object__close(trace->bpf_obj);
3604         trace->bpf_obj = NULL;
3605 }
3606 #else // HAVE_LIBBPF_SUPPORT
3607 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3608                                                    const char *name __maybe_unused)
3609 {
3610         return NULL;
3611 }
3612
3613 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3614 {
3615 }
3616
3617 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3618 {
3619 }
3620
3621 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3622 {
3623         return 0;
3624 }
3625
3626 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3627 {
3628         return 0;
3629 }
3630
3631 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3632                                                             const char *name __maybe_unused)
3633 {
3634         return NULL;
3635 }
3636
3637 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3638 {
3639         return 0;
3640 }
3641
3642 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3643 {
3644 }
3645 #endif // HAVE_LIBBPF_SUPPORT
3646
3647 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3648 {
3649         struct evsel *evsel;
3650
3651         evlist__for_each_entry(trace->evlist, evsel) {
3652                 if (evsel == trace->syscalls.events.augmented ||
3653                     evsel->bpf_obj == trace->bpf_obj)
3654                         continue;
3655
3656                 return false;
3657         }
3658
3659         return true;
3660 }
3661
3662 static int trace__set_ev_qualifier_filter(struct trace *trace)
3663 {
3664         if (trace->syscalls.map)
3665                 return trace__set_ev_qualifier_bpf_filter(trace);
3666         if (trace->syscalls.events.sys_enter)
3667                 return trace__set_ev_qualifier_tp_filter(trace);
3668         return 0;
3669 }
3670
3671 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3672                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3673 {
3674         int err = 0;
3675 #ifdef HAVE_LIBBPF_SUPPORT
3676         bool value = true;
3677         int map_fd = bpf_map__fd(map);
3678         size_t i;
3679
3680         for (i = 0; i < npids; ++i) {
3681                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3682                 if (err)
3683                         break;
3684         }
3685 #endif
3686         return err;
3687 }
3688
3689 static int trace__set_filter_loop_pids(struct trace *trace)
3690 {
3691         unsigned int nr = 1, err;
3692         pid_t pids[32] = {
3693                 getpid(),
3694         };
3695         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3696
3697         while (thread && nr < ARRAY_SIZE(pids)) {
3698                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3699
3700                 if (parent == NULL)
3701                         break;
3702
3703                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3704                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3705                         pids[nr++] = parent->tid;
3706                         break;
3707                 }
3708                 thread = parent;
3709         }
3710
3711         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3712         if (!err && trace->filter_pids.map)
3713                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3714
3715         return err;
3716 }
3717
3718 static int trace__set_filter_pids(struct trace *trace)
3719 {
3720         int err = 0;
3721         /*
3722          * Better not use !target__has_task() here because we need to cover the
3723          * case where no threads were specified in the command line, but a
3724          * workload was, and in that case we will fill in the thread_map when
3725          * we fork the workload in evlist__prepare_workload.
3726          */
3727         if (trace->filter_pids.nr > 0) {
3728                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3729                                                     trace->filter_pids.entries);
3730                 if (!err && trace->filter_pids.map) {
3731                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3732                                                        trace->filter_pids.entries);
3733                 }
3734         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3735                 err = trace__set_filter_loop_pids(trace);
3736         }
3737
3738         return err;
3739 }
3740
3741 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3742 {
3743         struct evlist *evlist = trace->evlist;
3744         struct perf_sample sample;
3745         int err = evlist__parse_sample(evlist, event, &sample);
3746
3747         if (err)
3748                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3749         else
3750                 trace__handle_event(trace, event, &sample);
3751
3752         return 0;
3753 }
3754
3755 static int __trace__flush_events(struct trace *trace)
3756 {
3757         u64 first = ordered_events__first_time(&trace->oe.data);
3758         u64 flush = trace->oe.last - NSEC_PER_SEC;
3759
3760         /* Is there some thing to flush.. */
3761         if (first && first < flush)
3762                 return ordered_events__flush_time(&trace->oe.data, flush);
3763
3764         return 0;
3765 }
3766
3767 static int trace__flush_events(struct trace *trace)
3768 {
3769         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3770 }
3771
3772 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3773 {
3774         int err;
3775
3776         if (!trace->sort_events)
3777                 return __trace__deliver_event(trace, event);
3778
3779         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3780         if (err && err != -1)
3781                 return err;
3782
3783         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3784         if (err)
3785                 return err;
3786
3787         return trace__flush_events(trace);
3788 }
3789
3790 static int ordered_events__deliver_event(struct ordered_events *oe,
3791                                          struct ordered_event *event)
3792 {
3793         struct trace *trace = container_of(oe, struct trace, oe.data);
3794
3795         return __trace__deliver_event(trace, event->event);
3796 }
3797
3798 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3799 {
3800         struct tep_format_field *field;
3801         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3802
3803         if (evsel->tp_format == NULL || fmt == NULL)
3804                 return NULL;
3805
3806         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3807                 if (strcmp(field->name, arg) == 0)
3808                         return fmt;
3809
3810         return NULL;
3811 }
3812
3813 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3814 {
3815         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3816
3817         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3818                 char *right = tok + 1, *right_end;
3819
3820                 if (*right == '=')
3821                         ++right;
3822
3823                 while (isspace(*right))
3824                         ++right;
3825
3826                 if (*right == '\0')
3827                         break;
3828
3829                 while (!isalpha(*left))
3830                         if (++left == tok) {
3831                                 /*
3832                                  * Bail out, can't find the name of the argument that is being
3833                                  * used in the filter, let it try to set this filter, will fail later.
3834                                  */
3835                                 return 0;
3836                         }
3837
3838                 right_end = right + 1;
3839                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3840                         ++right_end;
3841
3842                 if (isalpha(*right)) {
3843                         struct syscall_arg_fmt *fmt;
3844                         int left_size = tok - left,
3845                             right_size = right_end - right;
3846                         char arg[128];
3847
3848                         while (isspace(left[left_size - 1]))
3849                                 --left_size;
3850
3851                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3852
3853                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3854                         if (fmt == NULL) {
3855                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3856                                        arg, evsel->name, evsel->filter);
3857                                 return -1;
3858                         }
3859
3860                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3861                                  arg, (int)(right - tok), tok, right_size, right);
3862
3863                         if (fmt->strtoul) {
3864                                 u64 val;
3865                                 struct syscall_arg syscall_arg = {
3866                                         .parm = fmt->parm,
3867                                 };
3868
3869                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3870                                         char *n, expansion[19];
3871                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3872                                         int expansion_offset = right - new_filter;
3873
3874                                         pr_debug("%s", expansion);
3875
3876                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3877                                                 pr_debug(" out of memory!\n");
3878                                                 free(new_filter);
3879                                                 return -1;
3880                                         }
3881                                         if (new_filter != evsel->filter)
3882                                                 free(new_filter);
3883                                         left = n + expansion_offset + expansion_lenght;
3884                                         new_filter = n;
3885                                 } else {
3886                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3887                                                right_size, right, arg, evsel->name, evsel->filter);
3888                                         return -1;
3889                                 }
3890                         } else {
3891                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3892                                        arg, evsel->name, evsel->filter);
3893                                 return -1;
3894                         }
3895
3896                         pr_debug("\n");
3897                 } else {
3898                         left = right_end;
3899                 }
3900         }
3901
3902         if (new_filter != evsel->filter) {
3903                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3904                 evsel__set_filter(evsel, new_filter);
3905                 free(new_filter);
3906         }
3907
3908         return 0;
3909 }
3910
3911 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3912 {
3913         struct evlist *evlist = trace->evlist;
3914         struct evsel *evsel;
3915
3916         evlist__for_each_entry(evlist, evsel) {
3917                 if (evsel->filter == NULL)
3918                         continue;
3919
3920                 if (trace__expand_filter(trace, evsel)) {
3921                         *err_evsel = evsel;
3922                         return -1;
3923                 }
3924         }
3925
3926         return 0;
3927 }
3928
3929 static int trace__run(struct trace *trace, int argc, const char **argv)
3930 {
3931         struct evlist *evlist = trace->evlist;
3932         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3933         int err = -1, i;
3934         unsigned long before;
3935         const bool forks = argc > 0;
3936         bool draining = false;
3937
3938         trace->live = true;
3939         signal(SIGCHLD, sig_handler);
3940
3941         if (!trace->raw_augmented_syscalls) {
3942                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3943                         goto out_error_raw_syscalls;
3944
3945                 if (trace->trace_syscalls)
3946                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3947         }
3948
3949         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3950                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3951                 if (pgfault_maj == NULL)
3952                         goto out_error_mem;
3953                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3954                 evlist__add(evlist, pgfault_maj);
3955         }
3956
3957         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3958                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3959                 if (pgfault_min == NULL)
3960                         goto out_error_mem;
3961                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3962                 evlist__add(evlist, pgfault_min);
3963         }
3964
3965         if (trace->sched &&
3966             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3967                 goto out_error_sched_stat_runtime;
3968         /*
3969          * If a global cgroup was set, apply it to all the events without an
3970          * explicit cgroup. I.e.:
3971          *
3972          *      trace -G A -e sched:*switch
3973          *
3974          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3975          * _and_ sched:sched_switch to the 'A' cgroup, while:
3976          *
3977          * trace -e sched:*switch -G A
3978          *
3979          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3980          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3981          * a cgroup (on the root cgroup, sys wide, etc).
3982          *
3983          * Multiple cgroups:
3984          *
3985          * trace -G A -e sched:*switch -G B
3986          *
3987          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3988          * to the 'B' cgroup.
3989          *
3990          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3991          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3992          */
3993         if (trace->cgroup)
3994                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3995
3996         err = evlist__create_maps(evlist, &trace->opts.target);
3997         if (err < 0) {
3998                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3999                 goto out_delete_evlist;
4000         }
4001
4002         err = trace__symbols_init(trace, evlist);
4003         if (err < 0) {
4004                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
4005                 goto out_delete_evlist;
4006         }
4007
4008         evlist__config(evlist, &trace->opts, &callchain_param);
4009
4010         if (forks) {
4011                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
4012                 if (err < 0) {
4013                         fprintf(trace->output, "Couldn't run the workload!\n");
4014                         goto out_delete_evlist;
4015                 }
4016         }
4017
4018         err = evlist__open(evlist);
4019         if (err < 0)
4020                 goto out_error_open;
4021
4022         err = bpf__apply_obj_config();
4023         if (err) {
4024                 char errbuf[BUFSIZ];
4025
4026                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
4027                 pr_err("ERROR: Apply config to BPF failed: %s\n",
4028                          errbuf);
4029                 goto out_error_open;
4030         }
4031
4032         err = trace__set_filter_pids(trace);
4033         if (err < 0)
4034                 goto out_error_mem;
4035
4036         if (trace->syscalls.map)
4037                 trace__init_syscalls_bpf_map(trace);
4038
4039         if (trace->syscalls.prog_array.sys_enter)
4040                 trace__init_syscalls_bpf_prog_array_maps(trace);
4041
4042         if (trace->ev_qualifier_ids.nr > 0) {
4043                 err = trace__set_ev_qualifier_filter(trace);
4044                 if (err < 0)
4045                         goto out_errno;
4046
4047                 if (trace->syscalls.events.sys_exit) {
4048                         pr_debug("event qualifier tracepoint filter: %s\n",
4049                                  trace->syscalls.events.sys_exit->filter);
4050                 }
4051         }
4052
4053         /*
4054          * If the "close" syscall is not traced, then we will not have the
4055          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
4056          * fd->pathname table and were ending up showing the last value set by
4057          * syscalls opening a pathname and associating it with a descriptor or
4058          * reading it from /proc/pid/fd/ in cases where that doesn't make
4059          * sense.
4060          *
4061          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
4062          *  not in use.
4063          */
4064         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4065
4066         err = trace__expand_filters(trace, &evsel);
4067         if (err)
4068                 goto out_delete_evlist;
4069         err = evlist__apply_filters(evlist, &evsel);
4070         if (err < 0)
4071                 goto out_error_apply_filters;
4072
4073         if (trace->dump.map)
4074                 bpf_map__fprintf(trace->dump.map, trace->output);
4075
4076         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4077         if (err < 0)
4078                 goto out_error_mmap;
4079
4080         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
4081                 evlist__enable(evlist);
4082
4083         if (forks)
4084                 evlist__start_workload(evlist);
4085
4086         if (trace->opts.initial_delay) {
4087                 usleep(trace->opts.initial_delay * 1000);
4088                 evlist__enable(evlist);
4089         }
4090
4091         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4092                                   evlist->core.threads->nr > 1 ||
4093                                   evlist__first(evlist)->core.attr.inherit;
4094
4095         /*
4096          * Now that we already used evsel->core.attr to ask the kernel to setup the
4097          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4098          * trace__resolve_callchain(), allowing per-event max-stack settings
4099          * to override an explicitly set --max-stack global setting.
4100          */
4101         evlist__for_each_entry(evlist, evsel) {
4102                 if (evsel__has_callchain(evsel) &&
4103                     evsel->core.attr.sample_max_stack == 0)
4104                         evsel->core.attr.sample_max_stack = trace->max_stack;
4105         }
4106 again:
4107         before = trace->nr_events;
4108
4109         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4110                 union perf_event *event;
4111                 struct mmap *md;
4112
4113                 md = &evlist->mmap[i];
4114                 if (perf_mmap__read_init(&md->core) < 0)
4115                         continue;
4116
4117                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4118                         ++trace->nr_events;
4119
4120                         err = trace__deliver_event(trace, event);
4121                         if (err)
4122                                 goto out_disable;
4123
4124                         perf_mmap__consume(&md->core);
4125
4126                         if (interrupted)
4127                                 goto out_disable;
4128
4129                         if (done && !draining) {
4130                                 evlist__disable(evlist);
4131                                 draining = true;
4132                         }
4133                 }
4134                 perf_mmap__read_done(&md->core);
4135         }
4136
4137         if (trace->nr_events == before) {
4138                 int timeout = done ? 100 : -1;
4139
4140                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4141                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4142                                 draining = true;
4143
4144                         goto again;
4145                 } else {
4146                         if (trace__flush_events(trace))
4147                                 goto out_disable;
4148                 }
4149         } else {
4150                 goto again;
4151         }
4152
4153 out_disable:
4154         thread__zput(trace->current);
4155
4156         evlist__disable(evlist);
4157
4158         if (trace->sort_events)
4159                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4160
4161         if (!err) {
4162                 if (trace->summary)
4163                         trace__fprintf_thread_summary(trace, trace->output);
4164
4165                 if (trace->show_tool_stats) {
4166                         fprintf(trace->output, "Stats:\n "
4167                                                " vfs_getname : %" PRIu64 "\n"
4168                                                " proc_getname: %" PRIu64 "\n",
4169                                 trace->stats.vfs_getname,
4170                                 trace->stats.proc_getname);
4171                 }
4172         }
4173
4174 out_delete_evlist:
4175         trace__symbols__exit(trace);
4176         evlist__free_syscall_tp_fields(evlist);
4177         evlist__delete(evlist);
4178         cgroup__put(trace->cgroup);
4179         trace->evlist = NULL;
4180         trace->live = false;
4181         return err;
4182 {
4183         char errbuf[BUFSIZ];
4184
4185 out_error_sched_stat_runtime:
4186         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4187         goto out_error;
4188
4189 out_error_raw_syscalls:
4190         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4191         goto out_error;
4192
4193 out_error_mmap:
4194         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4195         goto out_error;
4196
4197 out_error_open:
4198         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4199
4200 out_error:
4201         fprintf(trace->output, "%s\n", errbuf);
4202         goto out_delete_evlist;
4203
4204 out_error_apply_filters:
4205         fprintf(trace->output,
4206                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4207                 evsel->filter, evsel__name(evsel), errno,
4208                 str_error_r(errno, errbuf, sizeof(errbuf)));
4209         goto out_delete_evlist;
4210 }
4211 out_error_mem:
4212         fprintf(trace->output, "Not enough memory to run!\n");
4213         goto out_delete_evlist;
4214
4215 out_errno:
4216         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4217         goto out_delete_evlist;
4218 }
4219
4220 static int trace__replay(struct trace *trace)
4221 {
4222         const struct evsel_str_handler handlers[] = {
4223                 { "probe:vfs_getname",       trace__vfs_getname, },
4224         };
4225         struct perf_data data = {
4226                 .path  = input_name,
4227                 .mode  = PERF_DATA_MODE_READ,
4228                 .force = trace->force,
4229         };
4230         struct perf_session *session;
4231         struct evsel *evsel;
4232         int err = -1;
4233
4234         trace->tool.sample        = trace__process_sample;
4235         trace->tool.mmap          = perf_event__process_mmap;
4236         trace->tool.mmap2         = perf_event__process_mmap2;
4237         trace->tool.comm          = perf_event__process_comm;
4238         trace->tool.exit          = perf_event__process_exit;
4239         trace->tool.fork          = perf_event__process_fork;
4240         trace->tool.attr          = perf_event__process_attr;
4241         trace->tool.tracing_data  = perf_event__process_tracing_data;
4242         trace->tool.build_id      = perf_event__process_build_id;
4243         trace->tool.namespaces    = perf_event__process_namespaces;
4244
4245         trace->tool.ordered_events = true;
4246         trace->tool.ordering_requires_timestamps = true;
4247
4248         /* add tid to output */
4249         trace->multiple_threads = true;
4250
4251         session = perf_session__new(&data, &trace->tool);
4252         if (IS_ERR(session))
4253                 return PTR_ERR(session);
4254
4255         if (trace->opts.target.pid)
4256                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4257
4258         if (trace->opts.target.tid)
4259                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4260
4261         if (symbol__init(&session->header.env) < 0)
4262                 goto out;
4263
4264         trace->host = &session->machines.host;
4265
4266         err = perf_session__set_tracepoints_handlers(session, handlers);
4267         if (err)
4268                 goto out;
4269
4270         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4271         /* older kernels have syscalls tp versus raw_syscalls */
4272         if (evsel == NULL)
4273                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4274
4275         if (evsel &&
4276             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4277             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4278                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4279                 goto out;
4280         }
4281
4282         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4283         if (evsel == NULL)
4284                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4285         if (evsel &&
4286             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4287             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4288                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4289                 goto out;
4290         }
4291
4292         evlist__for_each_entry(session->evlist, evsel) {
4293                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4294                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4295                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4296                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4297                         evsel->handler = trace__pgfault;
4298         }
4299
4300         setup_pager();
4301
4302         err = perf_session__process_events(session);
4303         if (err)
4304                 pr_err("Failed to process events, error %d", err);
4305
4306         else if (trace->summary)
4307                 trace__fprintf_thread_summary(trace, trace->output);
4308
4309 out:
4310         perf_session__delete(session);
4311
4312         return err;
4313 }
4314
4315 static size_t trace__fprintf_threads_header(FILE *fp)
4316 {
4317         size_t printed;
4318
4319         printed  = fprintf(fp, "\n Summary of events:\n\n");
4320
4321         return printed;
4322 }
4323
4324 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4325         struct syscall_stats *stats;
4326         double               msecs;
4327         int                  syscall;
4328 )
4329 {
4330         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4331         struct syscall_stats *stats = source->priv;
4332
4333         entry->syscall = source->i;
4334         entry->stats   = stats;
4335         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4336 }
4337
4338 static size_t thread__dump_stats(struct thread_trace *ttrace,
4339                                  struct trace *trace, FILE *fp)
4340 {
4341         size_t printed = 0;
4342         struct syscall *sc;
4343         struct rb_node *nd;
4344         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4345
4346         if (syscall_stats == NULL)
4347                 return 0;
4348
4349         printed += fprintf(fp, "\n");
4350
4351         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4352         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4353         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4354
4355         resort_rb__for_each_entry(nd, syscall_stats) {
4356                 struct syscall_stats *stats = syscall_stats_entry->stats;
4357                 if (stats) {
4358                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4359                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4360                         double avg = avg_stats(&stats->stats);
4361                         double pct;
4362                         u64 n = (u64)stats->stats.n;
4363
4364                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4365                         avg /= NSEC_PER_MSEC;
4366
4367                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4368                         printed += fprintf(fp, "   %-15s", sc->name);
4369                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4370                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4371                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4372
4373                         if (trace->errno_summary && stats->nr_failures) {
4374                                 const char *arch_name = perf_env__arch(trace->host->env);
4375                                 int e;
4376
4377                                 for (e = 0; e < stats->max_errno; ++e) {
4378                                         if (stats->errnos[e] != 0)
4379                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4380                                 }
4381                         }
4382                 }
4383         }
4384
4385         resort_rb__delete(syscall_stats);
4386         printed += fprintf(fp, "\n\n");
4387
4388         return printed;
4389 }
4390
4391 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4392 {
4393         size_t printed = 0;
4394         struct thread_trace *ttrace = thread__priv(thread);
4395         double ratio;
4396
4397         if (ttrace == NULL)
4398                 return 0;
4399
4400         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4401
4402         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4403         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4404         printed += fprintf(fp, "%.1f%%", ratio);
4405         if (ttrace->pfmaj)
4406                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4407         if (ttrace->pfmin)
4408                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4409         if (trace->sched)
4410                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4411         else if (fputc('\n', fp) != EOF)
4412                 ++printed;
4413
4414         printed += thread__dump_stats(ttrace, trace, fp);
4415
4416         return printed;
4417 }
4418
4419 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4420 {
4421         return ttrace ? ttrace->nr_events : 0;
4422 }
4423
4424 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4425         struct thread *thread;
4426 )
4427 {
4428         entry->thread = rb_entry(nd, struct thread, rb_node);
4429 }
4430
4431 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4432 {
4433         size_t printed = trace__fprintf_threads_header(fp);
4434         struct rb_node *nd;
4435         int i;
4436
4437         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4438                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4439
4440                 if (threads == NULL) {
4441                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4442                         return 0;
4443                 }
4444
4445                 resort_rb__for_each_entry(nd, threads)
4446                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4447
4448                 resort_rb__delete(threads);
4449         }
4450         return printed;
4451 }
4452
4453 static int trace__set_duration(const struct option *opt, const char *str,
4454                                int unset __maybe_unused)
4455 {
4456         struct trace *trace = opt->value;
4457
4458         trace->duration_filter = atof(str);
4459         return 0;
4460 }
4461
4462 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4463                                               int unset __maybe_unused)
4464 {
4465         int ret = -1;
4466         size_t i;
4467         struct trace *trace = opt->value;
4468         /*
4469          * FIXME: introduce a intarray class, plain parse csv and create a
4470          * { int nr, int entries[] } struct...
4471          */
4472         struct intlist *list = intlist__new(str);
4473
4474         if (list == NULL)
4475                 return -1;
4476
4477         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4478         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4479
4480         if (trace->filter_pids.entries == NULL)
4481                 goto out;
4482
4483         trace->filter_pids.entries[0] = getpid();
4484
4485         for (i = 1; i < trace->filter_pids.nr; ++i)
4486                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4487
4488         intlist__delete(list);
4489         ret = 0;
4490 out:
4491         return ret;
4492 }
4493
4494 static int trace__open_output(struct trace *trace, const char *filename)
4495 {
4496         struct stat st;
4497
4498         if (!stat(filename, &st) && st.st_size) {
4499                 char oldname[PATH_MAX];
4500
4501                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4502                 unlink(oldname);
4503                 rename(filename, oldname);
4504         }
4505
4506         trace->output = fopen(filename, "w");
4507
4508         return trace->output == NULL ? -errno : 0;
4509 }
4510
4511 static int parse_pagefaults(const struct option *opt, const char *str,
4512                             int unset __maybe_unused)
4513 {
4514         int *trace_pgfaults = opt->value;
4515
4516         if (strcmp(str, "all") == 0)
4517                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4518         else if (strcmp(str, "maj") == 0)
4519                 *trace_pgfaults |= TRACE_PFMAJ;
4520         else if (strcmp(str, "min") == 0)
4521                 *trace_pgfaults |= TRACE_PFMIN;
4522         else
4523                 return -1;
4524
4525         return 0;
4526 }
4527
4528 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4529 {
4530         struct evsel *evsel;
4531
4532         evlist__for_each_entry(evlist, evsel) {
4533                 if (evsel->handler == NULL)
4534                         evsel->handler = handler;
4535         }
4536 }
4537
4538 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4539 {
4540         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4541
4542         if (fmt) {
4543                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4544
4545                 if (scfmt) {
4546                         int skip = 0;
4547
4548                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4549                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4550                                 ++skip;
4551
4552                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4553                 }
4554         }
4555 }
4556
4557 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4558 {
4559         struct evsel *evsel;
4560
4561         evlist__for_each_entry(evlist, evsel) {
4562                 if (evsel->priv || !evsel->tp_format)
4563                         continue;
4564
4565                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4566                         evsel__init_tp_arg_scnprintf(evsel);
4567                         continue;
4568                 }
4569
4570                 if (evsel__init_syscall_tp(evsel))
4571                         return -1;
4572
4573                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4574                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4575
4576                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4577                                 return -1;
4578
4579                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4580                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4581                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4582
4583                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4584                                 return -1;
4585
4586                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4587                 }
4588         }
4589
4590         return 0;
4591 }
4592
4593 /*
4594  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4595  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4596  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4597  *
4598  * It'd be better to introduce a parse_options() variant that would return a
4599  * list with the terms it didn't match to an event...
4600  */
4601 static int trace__parse_events_option(const struct option *opt, const char *str,
4602                                       int unset __maybe_unused)
4603 {
4604         struct trace *trace = (struct trace *)opt->value;
4605         const char *s = str;
4606         char *sep = NULL, *lists[2] = { NULL, NULL, };
4607         int len = strlen(str) + 1, err = -1, list, idx;
4608         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4609         char group_name[PATH_MAX];
4610         struct syscall_fmt *fmt;
4611
4612         if (strace_groups_dir == NULL)
4613                 return -1;
4614
4615         if (*s == '!') {
4616                 ++s;
4617                 trace->not_ev_qualifier = true;
4618         }
4619
4620         while (1) {
4621                 if ((sep = strchr(s, ',')) != NULL)
4622                         *sep = '\0';
4623
4624                 list = 0;
4625                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4626                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4627                         list = 1;
4628                         goto do_concat;
4629                 }
4630
4631                 fmt = syscall_fmt__find_by_alias(s);
4632                 if (fmt != NULL) {
4633                         list = 1;
4634                         s = fmt->name;
4635                 } else {
4636                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4637                         if (access(group_name, R_OK) == 0)
4638                                 list = 1;
4639                 }
4640 do_concat:
4641                 if (lists[list]) {
4642                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4643                 } else {
4644                         lists[list] = malloc(len);
4645                         if (lists[list] == NULL)
4646                                 goto out;
4647                         strcpy(lists[list], s);
4648                 }
4649
4650                 if (!sep)
4651                         break;
4652
4653                 *sep = ',';
4654                 s = sep + 1;
4655         }
4656
4657         if (lists[1] != NULL) {
4658                 struct strlist_config slist_config = {
4659                         .dirname = strace_groups_dir,
4660                 };
4661
4662                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4663                 if (trace->ev_qualifier == NULL) {
4664                         fputs("Not enough memory to parse event qualifier", trace->output);
4665                         goto out;
4666                 }
4667
4668                 if (trace__validate_ev_qualifier(trace))
4669                         goto out;
4670                 trace->trace_syscalls = true;
4671         }
4672
4673         err = 0;
4674
4675         if (lists[0]) {
4676                 struct option o = {
4677                         .value = &trace->evlist,
4678                 };
4679                 err = parse_events_option(&o, lists[0], 0);
4680         }
4681 out:
4682         free(strace_groups_dir);
4683         free(lists[0]);
4684         free(lists[1]);
4685         if (sep)
4686                 *sep = ',';
4687
4688         return err;
4689 }
4690
4691 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4692 {
4693         struct trace *trace = opt->value;
4694
4695         if (!list_empty(&trace->evlist->core.entries)) {
4696                 struct option o = {
4697                         .value = &trace->evlist,
4698                 };
4699                 return parse_cgroups(&o, str, unset);
4700         }
4701         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4702
4703         return 0;
4704 }
4705
4706 static int trace__config(const char *var, const char *value, void *arg)
4707 {
4708         struct trace *trace = arg;
4709         int err = 0;
4710
4711         if (!strcmp(var, "trace.add_events")) {
4712                 trace->perfconfig_events = strdup(value);
4713                 if (trace->perfconfig_events == NULL) {
4714                         pr_err("Not enough memory for %s\n", "trace.add_events");
4715                         return -1;
4716                 }
4717         } else if (!strcmp(var, "trace.show_timestamp")) {
4718                 trace->show_tstamp = perf_config_bool(var, value);
4719         } else if (!strcmp(var, "trace.show_duration")) {
4720                 trace->show_duration = perf_config_bool(var, value);
4721         } else if (!strcmp(var, "trace.show_arg_names")) {
4722                 trace->show_arg_names = perf_config_bool(var, value);
4723                 if (!trace->show_arg_names)
4724                         trace->show_zeros = true;
4725         } else if (!strcmp(var, "trace.show_zeros")) {
4726                 bool new_show_zeros = perf_config_bool(var, value);
4727                 if (!trace->show_arg_names && !new_show_zeros) {
4728                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4729                         goto out;
4730                 }
4731                 trace->show_zeros = new_show_zeros;
4732         } else if (!strcmp(var, "trace.show_prefix")) {
4733                 trace->show_string_prefix = perf_config_bool(var, value);
4734         } else if (!strcmp(var, "trace.no_inherit")) {
4735                 trace->opts.no_inherit = perf_config_bool(var, value);
4736         } else if (!strcmp(var, "trace.args_alignment")) {
4737                 int args_alignment = 0;
4738                 if (perf_config_int(&args_alignment, var, value) == 0)
4739                         trace->args_alignment = args_alignment;
4740         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4741                 if (strcasecmp(value, "libtraceevent") == 0)
4742                         trace->libtraceevent_print = true;
4743                 else if (strcasecmp(value, "libbeauty") == 0)
4744                         trace->libtraceevent_print = false;
4745         }
4746 out:
4747         return err;
4748 }
4749
4750 static void trace__exit(struct trace *trace)
4751 {
4752         int i;
4753
4754         strlist__delete(trace->ev_qualifier);
4755         free(trace->ev_qualifier_ids.entries);
4756         if (trace->syscalls.table) {
4757                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4758                         syscall__exit(&trace->syscalls.table[i]);
4759                 free(trace->syscalls.table);
4760         }
4761         syscalltbl__delete(trace->sctbl);
4762         zfree(&trace->perfconfig_events);
4763 }
4764
4765 int cmd_trace(int argc, const char **argv)
4766 {
4767         const char *trace_usage[] = {
4768                 "perf trace [<options>] [<command>]",
4769                 "perf trace [<options>] -- <command> [<options>]",
4770                 "perf trace record [<options>] [<command>]",
4771                 "perf trace record [<options>] -- <command> [<options>]",
4772                 NULL
4773         };
4774         struct trace trace = {
4775                 .opts = {
4776                         .target = {
4777                                 .uid       = UINT_MAX,
4778                                 .uses_mmap = true,
4779                         },
4780                         .user_freq     = UINT_MAX,
4781                         .user_interval = ULLONG_MAX,
4782                         .no_buffering  = true,
4783                         .mmap_pages    = UINT_MAX,
4784                 },
4785                 .output = stderr,
4786                 .show_comm = true,
4787                 .show_tstamp = true,
4788                 .show_duration = true,
4789                 .show_arg_names = true,
4790                 .args_alignment = 70,
4791                 .trace_syscalls = false,
4792                 .kernel_syscallchains = false,
4793                 .max_stack = UINT_MAX,
4794                 .max_events = ULONG_MAX,
4795         };
4796         const char *map_dump_str = NULL;
4797         const char *output_name = NULL;
4798         const struct option trace_options[] = {
4799         OPT_CALLBACK('e', "event", &trace, "event",
4800                      "event/syscall selector. use 'perf list' to list available events",
4801                      trace__parse_events_option),
4802         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4803                      "event filter", parse_filter),
4804         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4805                     "show the thread COMM next to its id"),
4806         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4807         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4808                      trace__parse_events_option),
4809         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4810         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4811         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4812                     "trace events on existing process id"),
4813         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4814                     "trace events on existing thread id"),
4815         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4816                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4817         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4818                     "system-wide collection from all CPUs"),
4819         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4820                     "list of cpus to monitor"),
4821         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4822                     "child tasks do not inherit counters"),
4823         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4824                      "number of mmap data pages", evlist__parse_mmap_pages),
4825         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4826                    "user to profile"),
4827         OPT_CALLBACK(0, "duration", &trace, "float",
4828                      "show only events with duration > N.M ms",
4829                      trace__set_duration),
4830 #ifdef HAVE_LIBBPF_SUPPORT
4831         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4832 #endif
4833         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4834         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4835         OPT_BOOLEAN('T', "time", &trace.full_time,
4836                     "Show full timestamp, not time relative to first start"),
4837         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4838                     "Show only syscalls that failed"),
4839         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4840                     "Show only syscall summary with statistics"),
4841         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4842                     "Show all syscalls and summary with statistics"),
4843         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4844                     "Show errno stats per syscall, use with -s or -S"),
4845         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4846                      "Trace pagefaults", parse_pagefaults, "maj"),
4847         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4848         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4849         OPT_CALLBACK(0, "call-graph", &trace.opts,
4850                      "record_mode[,record_size]", record_callchain_help,
4851                      &record_parse_callchain_opt),
4852         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4853                     "Use libtraceevent to print the tracepoint arguments."),
4854         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4855                     "Show the kernel callchains on the syscall exit path"),
4856         OPT_ULONG(0, "max-events", &trace.max_events,
4857                 "Set the maximum number of events to print, exit after that is reached. "),
4858         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4859                      "Set the minimum stack depth when parsing the callchain, "
4860                      "anything below the specified depth will be ignored."),
4861         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4862                      "Set the maximum stack depth when parsing the callchain, "
4863                      "anything beyond the specified depth will be ignored. "
4864                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4865         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4866                         "Sort batch of events before processing, use if getting out of order events"),
4867         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4868                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4869         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4870                         "per thread proc mmap processing timeout in ms"),
4871         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4872                      trace__parse_cgroups),
4873         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4874                      "ms to wait before starting measurement after program "
4875                      "start"),
4876         OPTS_EVSWITCH(&trace.evswitch),
4877         OPT_END()
4878         };
4879         bool __maybe_unused max_stack_user_set = true;
4880         bool mmap_pages_user_set = true;
4881         struct evsel *evsel;
4882         const char * const trace_subcommands[] = { "record", NULL };
4883         int err = -1;
4884         char bf[BUFSIZ];
4885
4886         signal(SIGSEGV, sighandler_dump_stack);
4887         signal(SIGFPE, sighandler_dump_stack);
4888         signal(SIGINT, sig_handler);
4889
4890         trace.evlist = evlist__new();
4891         trace.sctbl = syscalltbl__new();
4892
4893         if (trace.evlist == NULL || trace.sctbl == NULL) {
4894                 pr_err("Not enough memory to run!\n");
4895                 err = -ENOMEM;
4896                 goto out;
4897         }
4898
4899         /*
4900          * Parsing .perfconfig may entail creating a BPF event, that may need
4901          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4902          * is too small. This affects just this process, not touching the
4903          * global setting. If it fails we'll get something in 'perf trace -v'
4904          * to help diagnose the problem.
4905          */
4906         rlimit__bump_memlock();
4907
4908         err = perf_config(trace__config, &trace);
4909         if (err)
4910                 goto out;
4911
4912         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4913                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4914
4915         /*
4916          * Here we already passed thru trace__parse_events_option() and it has
4917          * already figured out if -e syscall_name, if not but if --event
4918          * foo:bar was used, the user is interested _just_ in those, say,
4919          * tracepoint events, not in the strace-like syscall-name-based mode.
4920          *
4921          * This is important because we need to check if strace-like mode is
4922          * needed to decided if we should filter out the eBPF
4923          * __augmented_syscalls__ code, if it is in the mix, say, via
4924          * .perfconfig trace.add_events, and filter those out.
4925          */
4926         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4927             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4928                 trace.trace_syscalls = true;
4929         }
4930         /*
4931          * Now that we have --verbose figured out, lets see if we need to parse
4932          * events from .perfconfig, so that if those events fail parsing, say some
4933          * BPF program fails, then we'll be able to use --verbose to see what went
4934          * wrong in more detail.
4935          */
4936         if (trace.perfconfig_events != NULL) {
4937                 struct parse_events_error parse_err;
4938
4939                 parse_events_error__init(&parse_err);
4940                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4941                 if (err)
4942                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4943                 parse_events_error__exit(&parse_err);
4944                 if (err)
4945                         goto out;
4946         }
4947
4948         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4949                 usage_with_options_msg(trace_usage, trace_options,
4950                                        "cgroup monitoring only available in system-wide mode");
4951         }
4952
4953         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4954         if (IS_ERR(evsel)) {
4955                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4956                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4957                 goto out;
4958         }
4959
4960         if (evsel) {
4961                 trace.syscalls.events.augmented = evsel;
4962
4963                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4964                 if (evsel == NULL) {
4965                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4966                         goto out;
4967                 }
4968
4969                 if (evsel->bpf_obj == NULL) {
4970                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4971                         goto out;
4972                 }
4973
4974                 trace.bpf_obj = evsel->bpf_obj;
4975
4976                 /*
4977                  * If we have _just_ the augmenter event but don't have a
4978                  * explicit --syscalls, then assume we want all strace-like
4979                  * syscalls:
4980                  */
4981                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4982                         trace.trace_syscalls = true;
4983                 /*
4984                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4985                  * strace-like syscall tracing is not set, then we need to trow
4986                  * away the augmenter, i.e. all the events that were created
4987                  * from that BPF object file.
4988                  *
4989                  * This is more to fix the current .perfconfig trace.add_events
4990                  * style of setting up the strace-like eBPF based syscall point
4991                  * payload augmenter.
4992                  *
4993                  * All this complexity will be avoided by adding an alternative
4994                  * to trace.add_events in the form of
4995                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4996                  * need it.
4997                  *
4998                  * .perfconfig trace.add_events is still useful if we want, for
4999                  * instance, have msr_write.msr in some .perfconfig profile based
5000                  * 'perf trace --config determinism.profile' mode, where for some
5001                  * particular goal/workload type we want a set of events and
5002                  * output mode (with timings, etc) instead of having to add
5003                  * all via the command line.
5004                  *
5005                  * Also --config to specify an alternate .perfconfig file needs
5006                  * to be implemented.
5007                  */
5008                 if (!trace.trace_syscalls) {
5009                         trace__delete_augmented_syscalls(&trace);
5010                 } else {
5011                         trace__set_bpf_map_filtered_pids(&trace);
5012                         trace__set_bpf_map_syscalls(&trace);
5013                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
5014                 }
5015         }
5016
5017         err = bpf__setup_stdout(trace.evlist);
5018         if (err) {
5019                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
5020                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
5021                 goto out;
5022         }
5023
5024         err = -1;
5025
5026         if (map_dump_str) {
5027                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
5028                 if (trace.dump.map == NULL) {
5029                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
5030                         goto out;
5031                 }
5032         }
5033
5034         if (trace.trace_pgfaults) {
5035                 trace.opts.sample_address = true;
5036                 trace.opts.sample_time = true;
5037         }
5038
5039         if (trace.opts.mmap_pages == UINT_MAX)
5040                 mmap_pages_user_set = false;
5041
5042         if (trace.max_stack == UINT_MAX) {
5043                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
5044                 max_stack_user_set = false;
5045         }
5046
5047 #ifdef HAVE_DWARF_UNWIND_SUPPORT
5048         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
5049                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
5050         }
5051 #endif
5052
5053         if (callchain_param.enabled) {
5054                 if (!mmap_pages_user_set && geteuid() == 0)
5055                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
5056
5057                 symbol_conf.use_callchain = true;
5058         }
5059
5060         if (trace.evlist->core.nr_entries > 0) {
5061                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5062                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5063                         perror("failed to set syscalls:* tracepoint fields");
5064                         goto out;
5065                 }
5066         }
5067
5068         if (trace.sort_events) {
5069                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5070                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5071         }
5072
5073         /*
5074          * If we are augmenting syscalls, then combine what we put in the
5075          * __augmented_syscalls__ BPF map with what is in the
5076          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5077          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5078          *
5079          * We'll switch to look at two BPF maps, one for sys_enter and the
5080          * other for sys_exit when we start augmenting the sys_exit paths with
5081          * buffers that are being copied from kernel to userspace, think 'read'
5082          * syscall.
5083          */
5084         if (trace.syscalls.events.augmented) {
5085                 evlist__for_each_entry(trace.evlist, evsel) {
5086                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5087
5088                         if (raw_syscalls_sys_exit) {
5089                                 trace.raw_augmented_syscalls = true;
5090                                 goto init_augmented_syscall_tp;
5091                         }
5092
5093                         if (trace.syscalls.events.augmented->priv == NULL &&
5094                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5095                                 struct evsel *augmented = trace.syscalls.events.augmented;
5096                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5097                                     evsel__init_augmented_syscall_tp_args(augmented))
5098                                         goto out;
5099                                 /*
5100                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5101                                  * Above we made sure we can get from the payload the tp fields
5102                                  * that we get from syscalls:sys_enter tracefs format file.
5103                                  */
5104                                 augmented->handler = trace__sys_enter;
5105                                 /*
5106                                  * Now we do the same for the *syscalls:sys_enter event so that
5107                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5108                                  * as not to filter it, then we'll handle it just like we would
5109                                  * for the BPF_OUTPUT one:
5110                                  */
5111                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5112                                     evsel__init_augmented_syscall_tp_args(evsel))
5113                                         goto out;
5114                                 evsel->handler = trace__sys_enter;
5115                         }
5116
5117                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5118                                 struct syscall_tp *sc;
5119 init_augmented_syscall_tp:
5120                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5121                                         goto out;
5122                                 sc = __evsel__syscall_tp(evsel);
5123                                 /*
5124                                  * For now with BPF raw_augmented we hook into
5125                                  * raw_syscalls:sys_enter and there we get all
5126                                  * 6 syscall args plus the tracepoint common
5127                                  * fields and the syscall_nr (another long).
5128                                  * So we check if that is the case and if so
5129                                  * don't look after the sc->args_size but
5130                                  * always after the full raw_syscalls:sys_enter
5131                                  * payload, which is fixed.
5132                                  *
5133                                  * We'll revisit this later to pass
5134                                  * s->args_size to the BPF augmenter (now
5135                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5136                                  * so that it copies only what we need for each
5137                                  * syscall, like what happens when we use
5138                                  * syscalls:sys_enter_NAME, so that we reduce
5139                                  * the kernel/userspace traffic to just what is
5140                                  * needed for each syscall.
5141                                  */
5142                                 if (trace.raw_augmented_syscalls)
5143                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5144                                 evsel__init_augmented_syscall_tp_ret(evsel);
5145                                 evsel->handler = trace__sys_exit;
5146                         }
5147                 }
5148         }
5149
5150         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5151                 return trace__record(&trace, argc-1, &argv[1]);
5152
5153         /* Using just --errno-summary will trigger --summary */
5154         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5155                 trace.summary_only = true;
5156
5157         /* summary_only implies summary option, but don't overwrite summary if set */
5158         if (trace.summary_only)
5159                 trace.summary = trace.summary_only;
5160
5161         if (output_name != NULL) {
5162                 err = trace__open_output(&trace, output_name);
5163                 if (err < 0) {
5164                         perror("failed to create output file");
5165                         goto out;
5166                 }
5167         }
5168
5169         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5170         if (err)
5171                 goto out_close;
5172
5173         err = target__validate(&trace.opts.target);
5174         if (err) {
5175                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5176                 fprintf(trace.output, "%s", bf);
5177                 goto out_close;
5178         }
5179
5180         err = target__parse_uid(&trace.opts.target);
5181         if (err) {
5182                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5183                 fprintf(trace.output, "%s", bf);
5184                 goto out_close;
5185         }
5186
5187         if (!argc && target__none(&trace.opts.target))
5188                 trace.opts.target.system_wide = true;
5189
5190         if (input_name)
5191                 err = trace__replay(&trace);
5192         else
5193                 err = trace__run(&trace, argc, argv);
5194
5195 out_close:
5196         if (output_name != NULL)
5197                 fclose(trace.output);
5198 out:
5199         trace__exit(&trace);
5200         return err;
5201 }